USENIX Security '21 Fall Accepted Papers

Preventing abuse of web services by bots is an increasingly important problem, as abusive activities grow in both volume and variety. CAPTCHAs are the most common way for thwarting bot activities. However, they are often ineffective against bots and frustrating for humans. In addition, some recent CAPTCHA techniques diminish user privacy. Meanwhile, client-side Trusted Execution Environments (TEEs) are becoming increasingly widespread (notably, ARM TrustZone and Intel SGX), allowing establishment of trust in a small part (trust anchor or TCB) of client-side hardware. This prompts the question: can a TEE help reduce (or remove entirely) user burden of solving CAPTCHAs?

In this paper, we design CACTI: CAPTCHA Avoidance via Client-side TEE Integration. Using client-side TEEs, CACTI allows legitimate clients to generate unforgeable rate-proofs demonstrating how frequently they have performed specific actions. These rate-proofs can be sent to web servers in lieu of solving CAPTCHAs. CACTI provides strong client privacy guarantees, since the information is only sent to the visited website and authenticated using a group signature scheme. Our evaluations show that overall latency of generating and verifying a CACTI rate-proof is less than 0.25 sec, while CACTI's bandwidth overhead is over 98% lower than that of current CAPTCHA systems.

Current fallback authentication mechanisms are unreliable (e.g., security questions are easy to guess) and need improvement. Social authentication shows promise as a novel form of fallback authentication. In this paper, we report the results of a four-week study that explored people's perceived willingness to use video chat as a form of social authentication. We investigated whether people's mood, location, and trust, and the presence of others affected perceived willingness to use video chat to authenticate. We found that participants who were alone, reported a more positive mood, and had more trust in others reported more willingness to use video chat as an authentication method. Participants also reported more willingness to help others to authenticate via video chat than to initiate a video chat authentication session themselves. Our results provide initial insights into human-computer interaction issues that could stem from using video chat as a fallback authentication method within a small social network of people (e.g., family members and close friends) who know each other well and trust each other.

Hypervisors are widely deployed by cloud computing providers to support virtual machines, but their growing complexity poses a security risk, as large codebases contain many vulnerabilities. We present SeKVM, a layered Linux KVM hypervisor architecture that has been formally verified on multiprocessor hardware. Using layers, we isolate KVM's trusted computing base into a small core such that only the core needs to be verified to ensure KVM's security guarantees. Using layers, we model hardware features at different levels of abstraction tailored to each layer of software. Lower hypervisor layers that configure and control hardware are verified using a novel machine model that includes multiprocessor memory management hardware such as multi-level shared page tables, tagged TLBs, and a coherent cache hierarchy with cache bypass support. Higher hypervisor layers that build on the lower layers are then verified using a more abstract and simplified model, taking advantage of layer encapsulation to reduce proof burden. Furthermore, layers provide modularity to reduce verification effort across multiple implementation versions. We have retrofitted and verified multiple versions of KVM on Arm multiprocessor hardware, proving the correctness of the implementations and that they contain no vulnerabilities that can affect KVM's security guarantees. Our work is the first machine-checked proof for a commodity hypervisor using multiprocessor memory management hardware. SeKVM requires only modest KVM modifications and incurs only modest performance overhead versus unmodified KVM on real application workloads.

We debut domain shadowing, a novel censorship evasion technique leveraging content delivery networks (CDNs). Domain shadowing exploits the fact that CDNs allow their customers to claim arbitrary domains as the back-end. By setting the frond-end of a CDN service as an allowed domain and the back-end a blocked one, a censored user can access resources of the blocked domain with all "indicators", including the connecting URL, the SNI of the TLS connection, and the Host header of the HTTP(S) request, appear to belong to the allowed domain. Furthermore, we demonstrate that domain shadowing can be proliferated by domain fronting, a censorship evasion technique popularly used a few years ago, making it even more difficult to block. Compared with existing censorship evasion solutions, domain shadowing is lightweight, incurs negligible overhead, and does not require dedicated third-party support. As a proof of concept, we implemented domain shadowing as a Firefox browser extension and presented its capability in circumventing censorship within a heavily censored country known by its strict censorship policies and advanced technologies.

End-host runtime monitors (e.g., CFI, system call IDS) flag processes in response to symptoms of a possible attack. Unfortunately, the symptom (e.g., invalid control transfer) may occur long after the root cause (e.g., buffer overflow), creating a gap whereby bug reports received by developers contain (at best) a snapshot of the process long after it executed the buggy instructions. To help system administrators provide developers with more concise reports, we propose ARCUS, an automated framework that performs root cause analysis over the execution flagged by the end-host monitor. ARCUS works by testing “what if” questions to detect vulnerable states, systematically localizing bugs to their concise root cause while finding additional enforceable checks at the program binary level to demonstrably block them. Using hardware-supported processor tracing, ARCUS decouples the cost of analysis from host performance.

We have implemented ARCUS and evaluated it on 31 vulnerabilities across 20 programs along with over 9,000 test cases from the RIPE and Juliet suites. ARCUS identifies the root cause of all tested exploits — with 0 false positives or negatives — and even finds 4 new 0-day vulnerabilities in traces averaging 4,000,000 basic blocks. ARCUS handles programs compiled from upwards of 810,000 lines of C/C++ code without needing concrete inputs or re-execution.

Localized adversarial patches aim to induce misclassification in machine learning models by arbitrarily modifying pixels within a restricted region of an image. Such attacks can be realized in the physical world by attaching the adversarial patch to the object to be misclassified, and defending against such attacks is an unsolved/open problem. In this paper, we propose a general defense framework called PatchGuard that can achieve high provable robustness while maintaining high clean accuracy against localized adversarial patches. The cornerstone of PatchGuard involves the use of CNNs with small receptive fields to impose a bound on the number of features corrupted by an adversarial patch. Given a bounded number of corrupted features, the problem of designing an adversarial patch defense reduces to that of designing a secure feature aggregation mechanism. Towards this end, we present our robust masking defense that robustly detects and masks corrupted features to recover the correct prediction. Notably, we can prove the robustness of our defense against any adversary within our threat model. Our extensive evaluation on ImageNet, ImageNette (a 10-class subset of ImageNet), and CIFAR-10 datasets demonstrates that our defense achieves state-of-the-art performance in terms of both provable robust accuracy and clean accuracy.

Miscreants from online underground economies regularly exploit website vulnerabilities and inject fraudulent content into victim web pages to promote illicit goods and services. Scalable detection of such promotional website defacements remains an open problem despite their prevalence in Black Hat Search Engine Optimization (SEO) campaigns. Adversaries often manage to inject content in a stealthy manner by obfuscating the description of illicit products and/or the presence of defacements to make them undetectable. In this paper, we design and implement DMoS—a Defacement Monitoring System which protects websites from promotional defacements at scale. Our design is based on two key observations: Firstly, for effective advertising, the obfuscated jargons of illicit goods or services need to be easily understood by their target customers (i.e., sharing similar shape or pronunciation). Secondly, to promote the underground business, the defacements are crafted to boost search engine ranking of the defaced web pages while trying to stay stealthy from the maintainers and legitimate users of the compromised websites. Leveraging these insights, we first follow the human convention and design a jargon normalization algorithm to map obfuscated jargons to their original forms. We then develop a tag embedding mechanism, which enables DMoS to focus more on those not-so-visually-obvious, yet site-ranking influential HTML tags (i.e., title, meta). Consequently, DMoS can reliably detect illicit content hidden in compromised web pages. In particular, we have deployed DMoS as a cloud-based monitoring service for a five-month trial run. It has analyzed more than 38 million web pages across 7000+ commercial Chinese websites and found defacements in 11% of these websites. It achieves a recall over 99% with a precision about 89%. While the original design of DMoS focuses on the detection of Chinese promotional defacements, we have extended the system and demonstrated its applicability for English website defacement detection via proof-of-concept experiments.

The Bluetooth standard is ubiquitously supported by computers, smartphones, and IoT devices. Due to its complexity, implementations require large codebases, which are prone to security vulnerabilities, such as the recently discovered BlueBorne and BadBluetooth attacks. While defined by the standard, most of the Bluetooth functionality, as defined by different Bluetooth profiles, is not required in the common usage scenarios.

Starting from this observation, we implement LIGHTBLUE, a framework performing automatic, profile-aware debloating of Bluetooth stacks, allowing users to automatically minimize their Bluetooth attack surface by removing unneeded Bluetooth features. L IGHT B LUE starts with a target Bluetooth application, detects the associated Bluetooth profiles, and applies a combination of control-flow and data-flow analysis to remove unused code within a Bluetooth host code. Furthermore, to debloat the Bluetooth firmware, LIGHTBLUE extracts the used Host Controller Interface (HCI) commands and patches the HCI dispatcher in the Bluetooth firmware automatically, so that the Bluetooth firmware avoids processing unneeded HCI commands.

We evaluate LIGHTBLUE on four different Bluetooth hosts and three different Bluetooth controllers. Our evaluation shows that LIGHTB LUE achieves between 32% and 50% code reduction in the Bluetooth host code and between 57% and 83% HCI command reduction in the Bluetooth firmware. This code reduction leads to the prevention of attacks responsible for 20 known CVEs, such as BlueBorne and BadBluetooth, while introducing no performance overhead and without affecting the behavior of the debloated application.

Apple controls one of the largest mobile ecosystems, with 1.5 billion active devices worldwide, and offers twelve proprietary wireless Continuity services. Previous works have unveiled several security and privacy issues in the involved protocols. These works extensively studied AirDrop while the coverage of the remaining vast Continuity service space is still low. To facilitate the cumbersome reverse-engineering process, we describe the first guide on how to approach a structured analysis of the involved protocols using several vantage points available on macOS. Also, we develop a toolkit to automate parts of this otherwise manual process. Based on this guide, we analyze the full protocol stacks involved in three Continuity services, in particular, Handoff (HO), Universal Clipboard (UC), and Wi-Fi Password Sharing (PWS). We discover several vulnerabilities spanning from Bluetooth Low Energy (BLE) advertisements to Apple's proprietary authentication protocols. These flaws allow for device tracking via HO's mDNS responses, a denial-of-service (DoS) attack on HO and UC, a DoS attack on PWS that prevents Wi-Fi password entry, and a machine-in-the-middle (MitM) attack on PWS that connects a target to an attacker-controlled Wi-Fi network. Our PoC implementations demonstrate that the attacks can be mounted using affordable off-the-shelf hardware ($20 micro:bit and a Wi-Fi card). Finally, we suggest practical mitigations and share our findings with Apple, who have started to release fixes through iOS and macOS updates.

The increasing ubiquity of low-cost wireless sensors has enabled users to easily deploy systems to remotely monitor and control their environments. However, this raises privacy concerns for third-party occupants, such as a hotel room guest who may be unaware of deployed clandestine sensors. Previous methods focused on specific modalities such as detecting cameras but do not provide a generalized and comprehensive method to capture arbitrary sensors which may be "spying" on a user. In this work, we propose SnoopDog, a framework to not only detect common Wi-Fi-based wireless sensors that are actively monitoring a user, but also classify and localize each device. SnoopDog works by establishing causality between patterns in observable wireless traffic and a trusted sensor in the same space, e.g., an inertial measurement unit (IMU) that captures a user's movement. Once causality is established, SnoopDog performs packet inspection to inform the user about the monitoring device. Finally, SnoopDog localizes the clandestine device in a 2D plane using a novel trial-based localization technique. We evaluated SnoopDog across several devices and various modalities and were able to detect causality for snooping devices 95.2% of the time and localize devices to a sufficiently reduced sub-space.

Machine learning models are prone to memorizing sensitive data, making them vulnerable to membership inference attacks in which an adversary aims to guess if an input sample was used to train the model. In this paper, we show that prior work on membership inference attacks may severely underestimate the privacy risks by relying solely on training custom neural network classifiers to perform attacks and focusing only on the aggregate results over data samples, such as the attack accuracy. To overcome these limitations, we first propose to benchmark membership inference privacy risks by improving existing non-neural network based inference attacks and proposing a new inference attack method based on a modification of prediction entropy. We propose to supplement existing neural network based attacks with our proposed benchmark attacks to effectively measure the privacy risks. We also propose benchmarks for defense mechanisms by accounting for adaptive adversaries with knowledge of the defense and also accounting for the trade-off between model accuracy and privacy risks. Using our benchmark attacks, we demonstrate that existing defense approaches against membership inference attacks are not as effective as previously reported.

Next, we introduce a new approach for fine-grained privacy analysis by formulating and deriving a new metric called the privacy risk score. Our privacy risk score metric measures an individual sample's likelihood of being a training member, which allows an adversary to identify samples with high privacy risks and perform membership inference attacks with high confidence. We propose to combine both existing aggregate privacy analysis and our proposed fine-grained privacy analysis for systematically measuring privacy risks. We experimentally validate the effectiveness of the privacy risk score metric and demonstrate that the distribution of privacy risk scores across individual samples is heterogeneous. Finally, we perform an in-depth investigation to understand why certain samples have high privacy risk scores, including correlations with model properties such as model sensitivity, generalization error, and feature embeddings. Our work emphasizes the importance of a systematic and rigorous evaluation of privacy risks of machine learning models. We publicly release our code at https://github.com/inspire-group/membership-inference-evaluation and our evaluation mechanisms have also been integrated in Google's TensorFlow Privacy library.

Despite the prevalence of data breaches, there is a limited understanding of individuals' awareness, perception, and responses to breaches that affect them. We provide novel insights into this topic through an online study (n=413) in which we presented participants with up to three data breaches that had exposed their email addresses and other personal information. Overall, 73% of participants were affected by at least one breach, 5.36 breaches on average. Many participants attributed the cause of being affected by a breach} to their poor email and security practices; only 14% correctly attributed the cause to external factors such as breached organizations and hackers. Participants were unaware of 74% of displayed breaches and expressed various emotions when learning about them. While some reported intending to take action, most participants believed the breach would not impact them. Our findings underline the need for user-friendly tools to improve consumers' resilience against breaches and accountability for breached organizations to provide more proactive post-breach communications and mitigations.

Internet-wide scanning is a commonly used research technique that has helped uncover real-world attacks, find cryptographic weaknesses, and understand both operator and miscreant behavior. Studies that employ scanning have largely assumed that services are hosted on their IANA-assigned ports, overlooking the study of services on unusual ports. In this work, we investigate where Internet services are deployed in practice and evaluate the security posture of services on unexpected ports. We show protocol deployment is more diffuse than previously believed and that protocols run on many additional ports beyond their primary IANA-assigned port. For example, only 3% of HTTP and 6% of TLS services run on ports 80 and 443, respectively. Services on non-standard ports are more likely to be insecure, which results in studies dramatically underestimating the security posture of Internet hosts. Building on our observations, we introduce LZR ("Laser"), a system that identifies 99% of identifiable unexpected services in five handshakes and dramatically reduces the time needed to perform application-layer scans on ports with few responsive expected services (e.g., 5500% speedup on 27017/MongoDB). We conclude with recommendations for future studies.

Many organizations need large amounts of high quality data for their applications, and one way to acquire such data is to combine datasets from multiple parties. Since these organizations often own sensitive data that cannot be shared in the clear with others due to policy regulation and business competition, there is increased interest in utilizing secure multi-party computation (MPC). MPC allows multiple parties to jointly compute a function without revealing their inputs to each other. We present Cerebro, an end-to-end collaborative learning platform that enables parties to compute learning tasks without sharing plaintext data. By taking an end-to-end approach to the system design, Cerebro allows multiple parties with complex economic relationships to safely collaborate on machine learning computation through the use of release policies and auditing, while also enabling users to achieve good performance without manually navigating the complex performance tradeoffs between MPC protocols.

The ability to execute code in an emulator is a fundamental part of modern vulnerability testing. Unfortunately, this poses a challenge for many embedded systems, where firmware expects to interact with hardware devices specific to the target. Getting embedded system firmware to run outside its native environment, termed rehosting, requires emulating these hardware devices with enough accuracy to convince the firmware that it is executing on the target hardware. However, full fidelity emulation of target devices (which requires considerable engineering effort) may not be necessary to boot the firmware to a point of interest for an analyst (for example, a point where fuzzer input can be injected). We hypothesized that, for the firmware to boot successfully, it is sufficient to emulate only the behavior expected by the firmware, and that this behavior could be inferred automatically.

To test this hypothesis, we developed and implemented Jetset, a system that uses symbolic execution to infer what behavior firmware expects from a target device. Jetset can generate devices models for hardware peripherals in C, allowing an analyst to boot the firmware in an emulator (e.g., QEMU). We successfully applied Jetset to thirteen distinct pieces of firmware together representing three architectures, three application domains (power grid, avionics, and consumer electronics), and five different operating systems. We also demonstrate how Jetset-assisted rehosting facilitates fuzz-testing, a common security analysis technique, on an avionics embedded system, in which we found a previously unknown privilege escalation vulnerability.

Today's microprocessors often rely on microcode updates to address issues such as security or functional patches. Unfortunately, microcode update flexibility opens up new attack vectors through malicious microcode alterations. Such attacks share many features with hardware Trojans and have similar devastating consequences for system security. However, due to microcode's opaque nature, little is known in the open literature about the capabilities and limitations of microcode Trojans.

We introduce the design of a microcoded RISC-V processor architecture together with a microcode development and evaluation environment. Even though microcode typically has almost complete control of the processor hardware, the design of meaningful microcode Trojans is not straightforward. This somewhat counter-intuitive insight is due to the lack of information at the hardware level about the semantics of executed software. In three security case studies we demonstrate how to overcome these issues and give insights on how to design meaningful microcode Trojans that undermine system security. To foster future research and applications, we publicly release our implementation and evaluation platform.

The security of modern electronic devices relies on secret keys stored on secure hardware modules as the root-of-trust (RoT). Extracting those keys would break the security of the entire system. As shown before, sophisticated side-channel analysis (SCA) attacks, using chip failure analysis (FA) techniques, can extract data from on-chip memory cells. However, since the chip's layout is unknown to the adversary in practice, secret key localization and reverse engineering are onerous tasks. Consequently, hardware vendors commonly believe that the ever-growing physical complexity of the integrated circuit (IC) designs can be a natural barrier against potential adversaries. In this work, we present a novel approach that can extract the secret key without any knowledge of the IC's layout, and independent from the employed memory technology as key storage. We automate the—traditionally very labor-intensive—reverse engineering and data extraction process. To that end, we demonstrate that black-box measurements captured using laser-assisted SCA techniques from a training device with known key can be used to profile the device for a later key prediction on other victim devices with unknown keys. To showcase the potential of our approach, we target keys on three different hardware platforms, which are utilized as RoT in different products.

Dataflow tracking with Dynamic Taint Analysis (DTA) is an important method in systems security with many applications, including exploit analysis, guided fuzzing, and side-channel information leak detection. However, DTA is fundamentally limited by the Boolean nature of taint labels, which provide no information about the significance of detected dataflows and lead to false positives/negatives on complex real world programs.

We introduce proximal gradient analysis (PGA), a novel, theoretically grounded approach that can track more accurate and fine-grained dataflow information. PGA uses proximal gradients, a generalization of gradients for non-differentiable functions, to precisely compose gradients over non-differentiable operations in programs. Composing gradients over programs eliminates many of the dataflow propagation errors that occur in DTA and provides richer information about how each measured dataflow effects a program.

We compare our prototype PGA implementation to three state of the art DTA implementations on 7 real-world programs. Our results show that PGA can improve the F1 accuracy of data flow tracking by up to 33% over taint tracking (20% on average) without introducing any significant overhead (< 5% on average). We further demonstrate the effectiveness of PGA by discovering 22 bugs (20 confirmed by developers) and 2 side-channel leaks, and identifying exploitable dataflows in 19 existing CVEs in the tested programs.

Causality analysis automates attack forensic and facilitates behavioral detection by associating causally related but temporally distant system events. Despite its proven usefulness, the analysis suffers from the innate big data challenge to store and process a colossal amount of system events that are constantly collected from hundreds of thousands of end-hosts in a realistic network. In addition, the effectiveness of the analysis to discover security breaches relies on the assumption that comprehensive historical events over a long span are stored. Hence, it is imminent to address the scalability issue in order to make causality analysis practical and applicable to the enterprise-level environment.

In this work, we present SEAL, a novel data compression approach for causality analysis. Based on information-theoretic observations on system event data, our approach achieves lossless compression and supports near real-time retrieval of historic events. In the compression step, the causality graph induced by the system logs is investigated, and abundant edge reduction potentials are explored. In the query step, for maximal speed, decompression is opportunistically executed. Experiments on two real-world datasets show that SEAL offers 2.63x and 12.94x data size reduction, respectively. Besides, 89% of the queries are faster on the compressed dataset than the uncompressed one, and SEAL returns exactly the same query results as the uncompressed data.

Controller Area Network (CAN) has established itself as the main internal communication medium for vehicles. However, recent works have shown that error handling makes CAN nodes vulnerable to certain attacks. In the light of such a threat, we systematically analyze CAN's error handling and fault confinement mechanism to investigate it for further vulnerabilities. In this paper, we develop CANOX, a testing tool that monitors the behavior of a CAN node under different bus and error conditions, and flags conditions that cause an unexpected node behavior. Using CANOX, we found three major undiscovered vulnerabilities in the CAN standard that could be exploited to launch a variety of attacks. Combining the three vulnerabilities, we construct the Scan-Then-Strike Attack (STS), a multi-staged attack in which an attacker with no previous knowledge of the vehicle's internals maps the vehicle's CAN bus, identifies a safety-critical ECU, swiftly silences it, and persistently prevents it from recovering. We validate the practicality of STS by evaluating it on a CAN bus testbed and a real vehicle.

We introduce the first microarchitectural side channel attacks that leverage contention on the CPU ring interconnect. There are two challenges that make it uniquely difficult to exploit this channel. First, little is known about the ring interconnect's functioning and architecture. Second, information that can be learned by an attacker through ring contention is noisy by nature and has coarse spatial granularity. To address the first challenge, we perform a thorough reverse engineering of the sophisticated protocols that handle communication on the ring interconnect. With this knowledge, we build a cross-core covert channel over the ring interconnect with a capacity of over 4 Mbps from a single thread, the largest to date for a cross-core channel not relying on shared memory. To address the second challenge, we leverage the fine-grained temporal patterns of ring contention to infer a victim program's secrets. We demonstrate our attack by extracting key bits from vulnerable EdDSA and RSA implementations, as well as inferring the precise timing of keystrokes typed by a victim user.

Credit and debit-card payments are typically authenticated with PINs. Once entered into a terminal, the PIN is sent as an encrypted PIN block across a payments network to the destination bank, which decrypts and verifies the PIN block. Each node in the payments network routes the PIN block to the next node by decrypting the block with its own key, and then re-encrypting the PIN block with the next node's key; nodes establish shared secret keys with their neighbors to do so. This decrypt-then-encrypt operation over PIN blocks is known as PIN translation, and it is currently performed in Hardware Security Modules (HSMs) to avoid possible PIN exposure. However, HSMs incur heavy acquisition and operational expenses.

Introduced at EUROCRYPT '98, proxy re-encryption (PRE) is a cryptographic primitive which can re-encrypt without exposing sensitive data. We perform an extensive study of PRE as applied to PIN translation, and show through formalization, security analysis, and an implementation study that PRE is a practical alternative to HSMs. With PRE, we eliminate the need for HSMs during re-encryption of a PIN, thus greatly reducing the number of HSMs needed by each participant in the payments ecosystem. Along the way we conduct practice-oriented PRE research, with novel theoretical contributions to resolve issues in comparing so-called honest re-encryption to chosen ciphertext PRE security, and a new efficient PRE scheme achieving a type of chosen ciphertext security.

Logic locking is a promising solution against emerging hardware security threats, which entails protecting a Boolean circuit using a “keying” mechanism. The latest and hitherto unbroken logic-locking techniques are based on the “corrupt-and-correct (CAC)” principle, offering provable security against input-output query attacks. However, it remains unclear whether these techniques are susceptible to structural attacks. This paper exploits the properties of integrated circuit (IC) design tools, also termed electronic design automation (EDA) tools, to undermine the security of the CAC techniques. Our proposed attack can break all the CAC techniques, including the unbroken CACrem technique that 40+ hackers taking part in a competition for more than three months could not break. Our attack can break circuits processed with any EDA tools, which is alarming because, until now, none of the EDA tools can render a secure locking solution: logic locking cannot make use of the existing EDA tools. We also provide a security property to ensure resilience against structural attacks. The commonly-used circuits can satisfy this property but only in a few cases where they cannot even defeat brute-force; thus, questions arise on the use of these circuits as benchmarks to evaluate logic locking and other security techniques.

We are rethinking the decades-old design of the CAN bus by incorporating reactive defense capabilities in it. While its reliability and cost effectiveness turned CAN into the most widely used in-vehicle communication interface, its topology, physical layer and arbitration mechanism make it impossible to prevent certain types of adversarial activities on the bus. For example, DoS attacks cannot be stopped as the physical layer gives equal rights to all the connected ECUs and an adversary may exploit this by flooding the network with high priority frames or cause transmission errors which may move honest ECUs into the bus-off state. In response to this, we propose a reactive mechanism based on relays placed along the bus that will change the network topology in case of an attack, i.e., a moving target defense mechanism, allowing a bus guardian to filter and redirect legitimate traffic. We take care of physical properties of the bus and keep the 120 Ω load constant at the end of the lines whenever relays are triggered to modify the topology of the bus. We build a proof-of-concept implementation and test it in a laboratory setup with automotive-grade controllers that demonstrates its functionality over collected real-world in-vehicle traffic. Our experiments show that despite short term disturbances when the relays are triggered, the frame loss is effectively zero.

Protecting network protocols within an encrypted tunnel, using technologies such as Virtual Private Networks (VPNs), is increasingly important to millions of users needing solutions to evade censorship or protect their traffic against in/on-path observers/attackers. In this paper, we present a series of attacks from two threat models: an attacker that can inject spoofed packets into the network stack of a VPN client (called client-side), and an attacker that can spoof packets on the Internet and send them to a VPN server (called server-side). In both cases, we assume that the attacker is in/on-path, and can count encrypted bytes or packets over time. In both threat models, we demonstrate attacks to infer the existence of, interfere with, or inject data into TCP connections forwarded through the encrypted VPN tunnel. In the server-side threat model, we also demonstrate an attack to hijack tunneled DNS queries and completely remove the protections of the VPN tunnel. For the attacks presented in this paper, we (1) assess their feasibility in terms of packet rates and timing; (2) test their applicability against a broad range of VPN technologies, types, and vendors; and (3) consider practical issues with respect to real-world attacks. We followed an ethical disclosure process for all attacks presented in this paper. Client-side attacks were addressed with two CVEs and partially mitigated by a series of updates from some operating system and VPN client vendors. Server-side attacks have not been addressed and are still feasible with all operating systems and VPN servers that we tested.

Client-side CSRF is a new type of CSRF vulnerability where the adversary can trick the client-side JavaScript program to send a forged HTTP request to a vulnerable target site by modifying the program's input parameters. We have little-to-no knowledge of this new vulnerability, and exploratory security evaluations of JavaScript-based web applications are impeded by the scarcity of reliable and scalable testing techniques. This paper presents JAW, a framework that enables the analysis of modern web applications against client-side CSRF leveraging declarative traversals on hybrid property graphs, a canonical, hybrid model for JavaScript programs. We use JAW to evaluate the prevalence of client-side CSRF vulnerabilities among all (i.e., 106) web applications from the Bitnami catalog, covering over 228M lines of JavaScript code. Our approach uncovers 12,701 forgeable client-side requests affecting 87 web applications in total. For 203 forgeable requests, we successfully created client-side CSRF exploits against seven web applications that can execute arbitrary server-side state-changing operations or enable cross-site scripting and SQL injection, that are not reachable via the classical attack vectors. Finally, we analyzed the forgeable requests and identified 25 request templates, highlighting the fields that can be manipulated and the type of manipulation.

Recent years have seen the development of phishing detection and identification approaches to defend against phishing attacks. Phishing detection solutions often report binary results, i.e., phishing or not, without any explanation. In contrast, phishing identification approaches identify phishing webpages by visually comparing webpages with predefined legitimate references and report phishing along with its target brand, thereby having explainable results. However, there are technical challenges in visual analyses that limit existing solutions from being effective (with high accuracy) and efficient (with low runtime overhead), to be put to practical use.

In this work, we design a hybrid deep learning system, Phishpedia, to address two prominent technical challenges in phishing identification, i.e., (i) accurate recognition of identity logos on webpage screenshots, and (ii) matching logo variants of the same brand. Phishpedia achieves both high accuracy and low runtime overhead. And very importantly, different from common approaches, Phishpedia does not require training on any phishing samples. We carry out extensive experiments using real phishing data; the results demonstrate that Phishpedia significantly outperforms baseline identification approaches (EMD, PhishZoo, and LogoSENSE) in accurately and efficiently identifying phishing pages. We also deployed Phishpedia with CertStream service and discovered 1704 new real phishing websites within 30 days, way more than other solutions; moreover, 1113 of them are not reported by any engines in VirusTotal.

The traditional design principle for Internet protocols indicates: "Be strict when sending and tolerant when receiving" [RFC1958], and DNS is no exception to this. The transparency of DNS in handling the DNS records, also standardised specifically for DNS [RFC3597], is one of the key features that made it such a popular platform facilitating a constantly increasing number of new applications. An application simply creates a new DNS record and can instantly start distributing it over DNS without requiring any changes to the DNS servers and platforms. Our Internet wide study confirms that more than 1.3M (96% of tested) open DNS resolvers are standard compliant and treat DNS records transparently.

In this work we show that this `transparency' introduces a severe vulnerability in the Internet: we demonstrate a new method to launch string injection attacks by encoding malicious payloads into DNS records. We show how to weaponise such DNS records to attack popular applications. For instance, we apply string injection to launch a new type of DNS cache poisoning attack, which we evaluated against a population of open resolvers and found 105K to be vulnerable. Such cache poisoning cannot be prevented with common setups of DNSSEC. Our attacks apply to internal as well as to public services, for instance, we reveal that all eduroam services are vulnerable to our injection attacks, allowing us to launch exploits ranging from unauthorised access to eduroam networks to resource starvation. Depending on the application, our attacks cause system crashes, data corruption and leakage, degradation of security, and can introduce remote code execution and arbitrary errors.

In our evaluation of the attacks in the Internet we find that all the standard compliant open DNS resolvers we tested allow our injection attacks against applications and users on their networks.

Online social network (OSN) accounts are often more user-centric than other types of online accounts (e.g., email accounts) because they present a number of demographic attributes such as age, gender, location, and occupation. While these attributes allow for more meaningful online interactions, they can also be used by malicious parties to craft various types of abuse. To understand the effects of demographic attributes on attacker behavior in stolen social accounts, we devised a method to instrument and monitor such accounts. We then created, instrumented, and deployed more than 1000 Facebook accounts, and exposed them to criminals. Our results confirm that victim demographic traits indeed influence the way cybercriminals abuse their accounts. For example, we find that cybercriminals that access teen accounts write messages and posts more than the ones accessing adult accounts, and attackers that compromise male accounts perform disruptive activities such as changing some of their profile information more than the ones that access female accounts. This knowledge could potentially help online services develop new models to characterize benign and malicious activity across various demographic attributes, and thus automatically classify future activity.

Software-defined networking (SDN) has emerged as a flexible network architecture for central and programmatic control. Although SDN can improve network security oversight and policy enforcement, ensuring the security of SDN from sophisticated attacks is an ongoing challenge for practitioners. Existing network forensics tools attempt to identify and track such attacks, but holistic causal reasoning across control and data planes remains challenging.

We present PicoSDN, a provenance-informed causal observer for SDN attack analysis. PicoSDN leverages fine-grained data and execution partitioning techniques, as well as a unified control and data plane model, to allow practitioners to efficiently determine root causes of attacks and to make informed decisions on mitigating them. We implement PicoSDN on the popular ONOS SDN controller. Our evaluation across several attack case studies shows that PicoSDN is practical for the identification, analysis, and mitigation of SDN attacks.

Secure computation is a promising privacy enhancing technology, but it is often not scalable enough for data intensive applications. On the other hand, the use of sketches has gained popularity in data mining, because sketches often give rise to highly efficient and scalable sub-linear algorithms. It is natural to ask: what if we put secure computation and sketches together? We investigated the question and the findings are interesting: we can get security, we can get scalability, and somewhat unexpectedly, we can also get differential privacy—for free. Our study started from building a secure computation protocol based on the Flajolet-Martin (FM) sketches, for solving the Private Distributed Cardinality Estimation (PDCE) problem, which is a fundamental problem with applications ranging from crowd tracking to network monitoring. The state of art protocol for PDCE (Fenske et al. CCS '17) is computationally expensive and not scalable enough to cope with big data applications, which prompted us to design a better protocol. Our further analysis revealed that if the cardinality to be estimated is large enough, our protocol can achieve (ϵ,δ) differential privacy automatically, without requiring any additional manipulation of the output. The result signifies a new approach for achieving differential privacy that departs from the mainstream approach (i.e. adding noise to the result). Free differential privacy can be achieved because of two reasons: secure computation minimizes information leakage, and the intrinsic estimation variance of the FM sketch makes the output of our protocol uncertain. We further show that the result is not just theoretical: the minimal cardinality for differential privacy to hold is only 10² – 10⁴ for typical parameters.

The mitigation action against a malicious website may differ greatly depending on how that site is hosted. If it is hosted under a private apex domain, where all its subdomains and pages are under the apex domain owner's direct control, we could block at the apex domain level. If it is hosted under a public apex domain though (e.g., a web hosting service provider), it would be more appropriate to block at the subdomain level. Further, for the former case, the private apex domain may be legitimate but compromised, or may be attacker-generated, which, again, would warrant different mitigation actions: attacker-owned apex domains could be blocked permanently, while only temporarily for compromised ones.

In this paper, we study over eight hundred million VirusTotal (VT) URL scans from Aug. 1, 2019 to Nov. 18, 2019 and build the first content agnostic machine learning models to distinguish between the above mentioned different types of apex domains hosting malicious websites. Specifically, we first build a highly accurate model to distinguish between public and private apex domains. Then we build additional models to further distinguish compromised domains from attacker-owned ones. Utilizing our trained models, we conduct a large-scale study of the host domains of malicious websites. We observe that even though public apex domains are less than 1% of the apexes hosting malicious websites, they amount to a whopping 46.5% malicious web pages seen in VT URL feeds during our study period. 19.5% of these public malicious websites are compromised. Out of the remaining websites (53.5%), which are hosted on private apexes, we observe that attackers mostly compromise benign websites (65.6%) to launch their attacks, whereas only 34.4% of malicious websites are hosted on domains registered by attackers. Overall, we observe the concerning trend that the majority (81.7%) of malicious websites are hosted under apex domains that attackers do not own.

Our current defenses against IoT malware may not be adequate to remediate an IoT malware attack similar to the Mirai botnet. This work seeks to investigate this matter by systematically and empirically studying the lifecycle of IoT malware and comparing it with traditional malware that target desktop and mobile platforms. We present a large-scale measurement of more than 166K Linux-based IoT malware samples collected over a year. We compare our results with prior works by systematizing desktop and mobile malware studies into a novel framework and answering key questions about defense readiness. Based on our findings, we deduce that the required technology to defend against IoT malware is available, but we conclude that there are insufficient efforts in place to deal with a large-scale IoT malware infection breakout.

A large number of memory corruption vulnerabilities, e.g., heap overflow and use after free (UAF), could only be exploited in specific heap layouts via techniques like heap feng shui. To pave the way for automated exploit generation (AEG), automated heap layout manipulation is demanded.

In this paper, we present a novel solution MAZE to manipulate proof-of-concept (POC) samples' heap layouts. It first identifies heap layout primitives (i.e., input fragments or code snippets) available for users to manipulate the heap. Then, it applies a novel Dig & Fill algorithm, which models the problem as a Linear Diophantine Equation and solves it deterministically, to infer a primitive operation sequence that is able to generate target heap layout.

We implemented a prototype of MAZE based on the analysis engine S2E, and evaluated it on the PHP, Python and Perl interpreters and a set of CTF (capture the flag) programs, as well as a large micro-benchmark. Results showed that MAZE could generate expected heap layouts for over 90% of them.

An attacker can obtain a valid TLS certificate for a domain by hijacking communication between a certificate authority (CA) and a victim domain. Performing domain validation from multiple vantage points can defend against these attacks. We explore the design space of multi-vantage-point domain validation to achieve (1) security via sufficiently diverse vantage points, (2) performance by ensuring low latency and overhead in certificate issuance, (3) manageability by complying with CA/Browser forum requirements, and requiring minimal changes to CA operations, and (4) a low benign failure rate for legitimate requests. Our opensource implementation was deployed by the Let's Encrypt CA in February 2020, and has since secured the issuance of more than half a billion certificates during the first year of its deployment. Using real-world operational data from Let's Encrypt, we show that our approach has negligible latency and communication overhead, and a benign failure rate comparable to conventional designs with one vantage point. Finally, we evaluate the security improvements using a combination of ethically conducted real-world BGP hijacks, Internet-scale traceroute experiments, and a novel BGP simulation framework. We show that multi-vantage-point domain validation can thwart the vast majority of BGP attacks. Our work motivates the deployment of multi-vantage-point domain validation across the CA ecosystem to strengthen TLS certificate issuance and user privacy.

Password security hinges on an in-depth understanding of the techniques adopted by attackers. Unfortunately, real-world adversaries resort to pragmatic guessing strategies such as dictionary attacks that are inherently difficult to model in password security studies. In order to be representative of the actual threat, dictionary attacks must be thoughtfully configured and tuned. However, this process requires a domain-knowledge and expertise that cannot be easily replicated. The consequence of inaccurately calibrating dictionary attacks is the unreliability of password security analyses, impaired by a severe measurement bias.

In the present work, we introduce a new generation of dictionary attacks that is consistently more resilient to inadequate configurations. Requiring no supervision or domain-knowledge, this technique automatically approximates the advanced guessing strategies adopted by real-world attackers. To achieve this: (1) We use deep neural networks to model the proficiency of adversaries in building attack configurations. (2) Then, we introduce dynamic guessing strategies within dictionary attacks. These mimic experts' ability to adapt their guessing strategies on the fly by incorporating knowledge on their targets.

Our techniques enable more robust and sound password strength estimates within dictionary attacks, eventually reducing overestimation in modeling real-world threats in password security.

Known approaches for using decoy passwords (honeywords) to detect credential database breaches suffer from the need for a trusted component to recognize decoys when entered in login attempts, and from an attacker's ability to test stolen passwords at other sites to identify user-chosen passwords based on their reuse at those sites. Amnesia is a framework that resolves these difficulties. Amnesia requires no secret state to detect the entry of honeywords and additionally allows a site to monitor for the entry of its decoy passwords elsewhere. We quantify the benefits of Amnesia using probabilistic model checking and the practicality of this framework through measurements of a working implementation.

Despite their in-DRAM Target Row Refresh (TRR) mitigations, some of the most recent DDR4 modules are still vulnerable to many-sided Rowhammer bit flips. While these bit flips are exploitable from native code, triggering them in the browser from JavaScript faces three nontrivial challenges. First, given the lack of cache flushing instructions in JavaScript, existing eviction-based Rowhammer attacks are already slow for the older single- or double-sided variants and thus not always effective. With many-sided Rowhammer, mounting effective attacks is even more challenging, as it requires the eviction of many different aggressor addresses from the CPU caches. Second, the most effective many-sided variants, known as n-sided, require large physically-contiguous memory regions which are not available in JavaScript. Finally, as we show for the first time, eviction-based Rowhammer attacks require proper synchronization to bypass in-DRAM TRR mitigations.

Using a number of novel insights, we overcome these challenges to build SMASH (Synchronized MAny-Sided Hammering), a technique to succesfully trigger Rowhammer bit flips from JavaScript on modern DDR4 systems. To mount effective attacks, SMASH exploits high-level knowledge of cache replacement policies to generate optimal access patterns for eviction-based many-sided Rowhammer. To lift the requirement for large physically-contiguous memory regions, SMASH decomposes n-sided Rowhammer into multiple double-sided pairs, which we can identify using slice coloring. Finally, to bypass the in-DRAM TRR mitigations, SMASH carefully schedules cache hits and misses to successfully trigger synchronized many-sided Rowhammer bit flips. We showcase SMASH with an end-to-end JavaScript exploit which can fully compromise the Firefox browser in 15 minutes on average.

Deep learning has continued to show promising results for malware classification. However, to identify key malicious behaviors, malware analysts are still tasked with reverse engineering unknown malware binaries using static analysis tools, which can take hours. Although machine learning can be used to help identify important parts of a binary, supervised approaches are impractical due to the expense of acquiring a sufficiently large labeled dataset.

To increase the productivity of static (or manual) reverse engineering, we propose DeepReflect: a tool for localizing and identifying malware components within a malicious binary. To localize malware components, we use an unsupervised deep neural network in a novel way, and classify the components through a semi-supervised cluster analysis, where analysts incrementally provide labels during their daily work flow. The tool is practical since it requires no data labeling to train the localization model, and minimal/noninvasive labeling to train the classifier incrementally.

In our evaluation with five malware analysts on over 26k malware samples, we found that DeepReflect reduces the number of functions that an analyst needs to reverse engineer by 85% on average. Our approach also detects 80% of the malware components compared to 43% when using a signature-based tool (CAPA). Furthermore, DeepReflect performs better with our proposed autoencoder than SHAP (an AI explanation tool). This is significant because SHAP, a state-of-the-art method, requires a labeled dataset and autoencoders do not.

TLS is widely used to add confidentiality, authenticity and integrity to application layer protocols such as HTTP, SMTP, IMAP, POP3, and FTP. However, TLS does not bind a TCP connection to the intended application layer protocol. This allows a man-in-the-middle attacker to redirect TLS traffic to a different TLS service endpoint on another IP address and/or port. For example, if subdomains share a wildcard certificate, an attacker can redirect traffic from one subdomain to another, resulting in a valid TLS session. This breaks the authentication of TLS and cross-protocol attacks may be possible where the behavior of one service may compromise the security of the other at the application layer.

In this paper, we investigate cross-protocol attacks on TLS in general and conduct a systematic case study on web servers, redirecting HTTPS requests from a victim's web browser to SMTP, IMAP, POP3, and FTP servers. We show that in realistic scenarios, the attacker can extract session cookies and other private user data or execute arbitrary JavaScript in the context of the vulnerable web server, therefore bypassing TLS and web application security.

We evaluate the real-world attack surface of web browsers and widely-deployed email and FTP servers in lab experiments and with internet-wide scans. We find that 1.4M web servers are generally vulnerable to cross-protocol attacks, i.e., TLS application data confusion is possible. Of these, 114k web servers can be attacked using an exploitable application server. Finally, we discuss the effectiveness of TLS extensions such as Application Layer Protocol Negotiation (ALPN) and Server Name Indiciation (SNI) in mitigating these and other cross-protocol attacks.

One intriguing property of deep neural networks (DNNs) is their inherent vulnerability to backdoor attacks—a trojan model responds to trigger-embedded inputs in a highly predictable manner while functioning normally otherwise. Despite the plethora of prior work on DNNs for continuous data (e.g., images), the vulnerability of graph neural networks (GNNs) for discrete-structured data (e.g., graphs) is largely unexplored, which is highly concerning given their increasing use in security-sensitive domains.

To bridge this gap, we present GTA, the first backdoor attack on GNNs. Compared with prior work, GTA departs in significant ways: graph-oriented—it defines triggers as specific subgraphs, including both topological structures and descriptive features, entailing a large design spectrum for the adversary; input-tailored—it dynamically adapts triggers to individual graphs, thereby optimizing both attack effectiveness and evasiveness; downstream model-agnostic—it can be readily launched without knowledge regarding downstream models or fine-tuning strategies; and attack-extensible—it can be instantiated for both transductive (e.g., node classification) and inductive (e.g., graph classification) tasks, constituting severe threats for a range of security-critical applications. Through extensive evaluation using benchmark datasets and state-of-the-art models, we demonstrate the effectiveness of GTA. We further provide analytical justification for its effectiveness and discuss potential countermeasures, pointing to several promising research directions.

We describe Swivel, a new compiler framework for hardening WebAssembly (Wasm) against Spectre attacks. Outside the browser, Wasm has become a popular lightweight, in-process sandbox and is, for example, used in production to isolate different clients on edge clouds and function-as-a-service platforms. Unfortunately, Spectre attacks can bypass Wasm's isolation guarantees. Swivel hardens Wasm against this class of attacks by ensuring that potentially malicious code can neither use Spectre attacks to break out of the Wasm sandbox nor coerce victim code—another Wasm client or the embedding process—to leak secret data.

We describe two Swivel designs, a software-only approach that can be used on existing CPUs, and a hardware-assisted approach that uses extension available in Intel® 11th generation CPUs. For both, we evaluate a randomized approach that mitigates Spectre and a deterministic approach that eliminates Spectre altogether. Our randomized implementations impose under 10.3% overhead on the Wasm-compatible subset of SPEC 2006, while our deterministic implementations impose overheads between 3.3% and 240.2%. Though high on some benchmarks, Swivel's overhead is still between 9× and 36.3× smaller than existing defenses that rely on pipeline fences.

On-device machine learning (ML) is quickly gaining popularity among mobile apps. It allows offline model inference while preserving user privacy. However, ML models, considered as core intellectual properties of model owners, are now stored on billions of untrusted devices and subject to potential thefts. Leaked models can cause both severe financial loss and security consequences. This paper presents the first empirical study of ML model protection on mobile devices. Our study aims to answer three open questions with quantitative evidence: How widely is model protection used in apps? How robust are existing model protection techniques? What impacts can (stolen) models incur? To that end, we built a simple app analysis pipeline and analyzed 46,753 popular apps collected from the US and Chinese app markets. We identified 1,468 ML apps spanning all popular app categories. We found that, alarmingly, 41% of ML apps do not protect their models at all, which can be trivially stolen from app packages. Even for those apps that use model protection or encryption, we were able to extract the models from 66% of them via unsophisticated dynamic analysis techniques. The extracted models are mostly commercial products and used for face recognition, liveness detection, ID/bank card recognition, and malware detection. We quantitatively estimated the potential financial and security impact of a leaked model, which can amount to millions of dollars for different stakeholders. Our study reveals that on-device models are currently at high risk of being leaked; attackers are highly motivated to steal such models. Drawn from our large-scale study, we report our insights into this emerging security problem and discuss the technical challenges, hoping to inspire future research on robust and practical model protection for mobile devices.

There are various costs for attackers to manipulate the features of security classifiers. The costs are asymmetric across features and to the directions of changes, which cannot be precisely captured by existing cost models based on L_p-norm robustness. In this paper, we utilize such domain knowledge to increase the attack cost of evading classifiers, specifically, tree ensemble models that are widely used by security tasks. We propose a new cost modeling method to capture the feature manipulation cost as constraint, and then we integrate the cost-driven constraint into the node construction process to train robust tree ensembles. During the training process, we use the constraint to find data points that are likely to be perturbed given the feature manipulation cost, and we use a new robust training algorithm to optimize the quality of the trees. Our cost-aware training method can be applied to different types of tree ensembles, including gradient boosted decision trees and random forest models. Using Twitter spam detection as the case study, our evaluation results show that we can increase the attack cost by 10.6X compared to the baseline. Moreover, our robust training method using cost-driven constraint can achieve higher accuracy, lower false positive rate, and stronger cost-aware robustness than the state-of-the-art training method using L_∞-norm cost model. Our code is available at https://github.com/surrealyz/growtrees.

Safeguarding blockchain peer-to-peer (P2P) networks is more critical than ever in light of recent network attacks. Bitcoin has been successfully handling traditional Sybil and eclipse attacks; however, a recent Erebus attack [Tran et al. IEEE S&P'20] shows that effectively eclipsing a Bitcoin node is possible when the attack is combined with a network-Sybil capability; i.e., a malicious transit network can create millions or more Sybil identities. Given the immediate availability and stealthiness of the Erebus attack, Bitcoin Core has quickly implemented a few simple protocol/parameter changes to mitigate it. Our large-scale evaluations of these quick patches and three similar carefully-designed protocol tweaks confirm that, unfortunately, no simple solution can effectively handle the attack. This paper focuses on a more fundamental solution called routing-aware peering (or RAP), a proven silver bullet in detecting and circumventing similar network adversaries in other P2P networks. However, we show that, contrary to our expectation, preventing the Erebus attacks with RAP is only wishful thinking. We discover that Erebus adversaries can exploit a tiny portion of route inference errors in any RAP implementations, which gives an asymmetric advantage to the network adversaries and renders all RAP approaches ineffective. To that end, we propose an integrated defense framework that composes the available simple protocol tweaks and RAP implementation. In particular, we show that a highly customizable defense profile is required for individual Bitcoin nodes because RAP's efficacy depends significantly on where a Bitcoin node is located on the Internet topology. We present an algorithm that outputs a custom optimal defense profile that prevents most of Erebus attacks from the top-100 large transit networks.

Many control-flow integrity (CFI) solutions have been proposed to protect indirect control transfers (ICT), including C++ virtual calls. Assessing the security guarantees of these defenses is thus important but hard. In practice, for a (strong) defense, it usually requires abundant manual efforts to assess whether it could be bypassed, when given a specific (weak) vulnerability. Existing automated exploit generation solutions, which are proposed to assess the exploitability of vulnerabilities, have not addressed this issue yet.

In this paper, we point out that a wide range of virtual call protections, which do not break the C++ ABI (application binary interface), are vulnerable to an advanced attack COOPLUS, even if the given vulnerabilities are weak. Then, we present a solution VScape to assess the effectiveness of virtual call protections against this attack. We developed a prototype of VScape, and utilized it to assess 11 CFI solutions and 14 C++ applications (including Firefox and PyQt) with known vulnerabilities. Results showed that real-world applications have a large set of exploitable virtual calls, and VScape could be utilized to generate working exploits to bypass deployed defenses via weak vulnerabilities.

IoT devices have brought invaluable convenience to our daily life. However, their pervasiveness also amplifies the impact of security vulnerabilities. Many popular vulnerabilities of embedded systems reside in their vulnerable web services. Unfortunately, existing vulnerability detection methods cannot effectively nor efficiently analyze such web services: they either introduce heavy execution overheads or have many false positives and false negatives.

In this paper, we propose a novel static taint checking solution, SaTC, to effectively detect security vulnerabilities in web services provided by embedded devices. Our key insight is that, string literals on web interfaces are commonly shared between front-end files and back-end binaries to encode user input. We thus extract such common keywords from the front-end, and use them to locate reference points in the back-end, which indicate the input entry. Then, we apply targeted data-flow analysis to accurately detect dangerous uses of the untrusted user input. We implemented a prototype of SaTC and evaluated it on 39 embedded system firmwares from six popular vendors. SaTC discovered 33 unknown bugs, of which 30 are confirmed by CVE/CNVD/PSV. Compared to the state-of-the-art tool KARONTE, SaTC found significantly more bugs on the test set. It shows that, SaTC is effective in discovering bugs in embedded systems.

Recent years have witnessed the rise of security risks of libraries integrated in mobile apps, which are reported to steal private user data from the host apps and the app backend servers. Their security implications, however, have never been fully understood. In our research, we brought to light a new attack vector long been ignored yet with serious privacy impacts – malicious libraries strategically target other vendors'SDKs integrated in the same host app to harvest private user data (e.g., Facebook's user profile). Using a methodology that incorporates semantic analysis on an SDK's Terms of Services (ToS, which describes restricted data access and sharing policies) and code analysis on cross-library interactions, we were able to investigate 1.3 million Google Play apps and the ToSes from 40 highly-popular SDKs, leading to the discovery of 42 distinct libraries stealthily harvesting data from 16 popular SDKs, which affect more than 19K apps with a total of 9 billion downloads. Our study further sheds light on the underground ecosystem behind such library-based data harvesting (e.g., monetary incentives for SDK integration), their unique strategies (e.g., hiding data in crash reports and using C2 server to schedule data exfiltration) and significant impacts.

This work introduces a novel four-party honest-majority MPC protocol with active security that achieves comparable efficiency to equivalent protocols in the same setting, while having a much simpler design and not relying on function-dependent preprocessing. Our initial protocol satisfies security with abort, but we present some extensions to achieve guaranteed output delivery. Unlike previous works, we do not achieve this by delegating the computation to one single party that is identified to be honest, which is likely to hinder the adoption of these technologies as it centralizes sensitive data. Instead, our novel approach guarantees termination of the protocol while ensuring that no single party (honest or corrupt) learns anything beyond the output.

We implement our four-party protocol with abort in the MP-SPDZ framework for multi-party computation and benchmark multiple applications like MNIST classification training and ImageNet inference. Our results show that our four-party protocol performs similarly to an efficient honest-majority three-party protocol that only provides semi-honest/passive security, which suggests that adding a fourth party can be an effective method to achieve active security without harming performance.

The EOSIO blockchain, one of the representative Delegated Proof-of-Stake (DPoS) blockchain platforms, has grown rapidly recently. Meanwhile, a number of vulnerabilities and high-profile attacks against top EOSIO DApps and their smart contracts have also been discovered and observed in the wild, resulting in serious financial damages. Most of the EOSIO smart contracts are not open-sourced and typically compiled to WebAssembly (Wasm) bytecode, thus making it challenging to analyze and detect the presence of possible vulnerabilities. In this paper, we propose EOSAFE, the first static analysis framework that can be used to automatically detect vulnerabilities in EOSIO smart contracts at the bytecode level. Our framework includes a practical symbolic execution engine for Wasm, a customized library emulator for EOSIO smart contracts, and four heuristic-driven detectors to identify the presence of the four most popular vulnerabilities in EOSIO smart contracts. Experiments have shown that EOSAFE achieves promising results in detecting vulnerabilities, with an F1-measure of 98%. We have applied EOSAFE to all active 53,666 smart contracts in the ecosystem (as of November 15, 2019). Our results show that over 25% of the smart contracts are labeled vulnerable. We further analyze possible exploitation attempts on these vulnerable smart contracts and identify 48 in-the-wild attacks (27 of them have been confirmed by DApp developers), which have resulted in financial loss of at least 1.7 million USD.

This paper presents our study on an emerging paradigm of payment service that allows individual merchants to leverage the personal transfer service in third-party platforms to support commercial transactions. This is made possible by leveraging an additional order management system, collectively named Personal Payment System (PPS). To gain a better understanding of these emerging systems, we conducted a systematic study on 35 PPSs covering over 11740 merchant clients supporting more than 20 million customers. By examining the documentation, available source codes, and demos, we extracted a common abstracted model for PPS and discovered seven categories of vulnerabilities in the existing personal payment protocol design and system implementation. It is alarming that all PPSs under study have at least one vulnerability. To further dissect these potential weaknesses, we present the corresponding attack methods to exploit the discovered vulnerabilities. To validate our proposed attacks, we conducted four successful real attacks to illustrate the severe consequences. We have responsibly disclosed the newly discovered vulnerabilities, with some patched after our reporting.

The 3GPP consortium has published the Authentication and Key Agreement protocol for the 5th generation (5G) mobile communication system (i.e., 5G-AKA) by Technical Specification (TS) 33.501. It introduces public key encryption to conceal the so-called SUPIs so as to enhance mobile users' privacy. However, 5G-AKA is only privacy-preserving at the presence of passive attackers, and is still vulnerable to the linkability attacks from active attackers. An active attacker can track target mobile phones via performing these attacks, which puts the privacy of users at risk.

It is resistant to linkability attacks performed by active attackers, and is compatible with the SIM cards and currently deployed Serving Networks (SNs). In particular, we first conduct an analysis on the known linkability attacks in 5G-AKA, and find out a root cause of all attacks. Then, we design a counter-measure with the inherent key encapsulation mechanism of ECIES (i.e., ECIES-KEM), and use the shared key established by ECIES-KEM to encrypt the challenges sent by a Home Network (HN). With this measure, a target User Equipment (UE) who receives a message replayed from its previously attended sessions behaves as non-target UEs, which prevents the attacker from distinguishing the UE by linking it with its previous sessions. Moreover, 5G-AKA′ does not raise additional bandwidth cost, and only introduces limited additional time costs from 0.02% to 0.03%. Finally, we use a state-of-the-art formal verification tool, Tamarin prover, to prove that 5G-AKA′ achieves the desired security goals of privacy, authentication and secrecy.

Accessibility features of Android are crucial in assisting people with disabilities or impairment to navigate their devices. However, the same, powerful features are commonly misused by shady apps for malevolent purposes, such as stealing data from other apps. Unfortunately, existing defenses do not allow apps to protect themselves and at the same time to be fully inclusive to users with accessibility needs.

To enhance the privacy protection of the user while preserving the accessibility features for assistive apps, we introduce an extension to Android's accessibility framework. Our design is based on a study of how accessibility features are used in 95 existing accessibility apps of different types (malware, utility, and a11y). Based on those insights, we propose to model the usage of the accessibility framework as a pipeline of code modules, which are all sandboxed on the system-side. By policing the data flows of those modules, we achieve more fine-grained control over the access to accessibility features and the way they are used in apps, allowing a balance between accessibility functionality for dependent users and reduced privacy risks. We demonstrate the feasibility of our solution by migrating two real-world apps to our privacy-enhanced accessibility framework.

CAPTCHA is an effective mechanism for protecting computers from malicious bots. With the development of deep learning techniques, current mainstream text-based CAPTCHAs have been proven to be insecure. Therefore, a major effort has been directed toward developing image-based CAPTCHAs, and image-based visual reasoning is emerging as a new direction of such development. Recently, Tencent deployed the Visual Turing Test (VTT) CAPTCHA. This appears to have been the first application of a visual reasoning scheme. Subsequently, other CAPTCHA service providers (Geetest, NetEase, Dingxiang, etc.) have proposed their own visual reasoning schemes to defend against bots. It is, therefore, natural to ask a fundamental question: are visual reasoning CAPTCHAs as secure as their designers expect? This paper presents the first attempt to solve visual reasoning CAPTCHAs. We implemented a holistic attack and a modular attack, which achieved overall success rates of 67.3% and 88.0% on VTT CAPTCHA, respectively. The results show that visual reasoning CAPTCHAs are not as secure as anticipated; this latest effort to use novel, hard AI problems for CAPTCHAs has not yet succeeded. Based on the lessons we learned from our attacks, we also offer some guidelines for designing visual CAPTCHAs with better security.

Neural-network classification is getting more pervasive. It captures data of the subjects to be classified, e.g., appearance for facial recognition, which is personal and often sensitive. Oblivious inference protects the data privacy of both the query and the model. However, it is not as fast and as accurate as its plaintext counterpart. A recent cryptographic solution Delphi (Usenix Security 2020) strives for low latency by using GPU on linear layers and replacing some non-linear units in the model at a price of accuracy. It can handle a query on CIFAR-100 with ~68% accuracy in 14s or ~66% accuracy in 2.6s.

We propose GForce, tackling the latency issue from the root causes instead of approximating non-linear computations. With the SWALP training approach (ICML 2019), we propose stochastic rounding and truncation (SRT) layers, which fuse quantization with dequantization between non-linear and linear layers and free us from floating-point operations for efficiency. They also ensure high accuracy while working over the severely-finite cryptographic field. We further propose a suite of GPU-friendly secure online/offline protocols for common operations, including comparison and wrap-around handling, which benefit non-linear layers, including our SRT.

With our two innovations, GForce supports VGG16, attaining ~73% accuracy over CIFAR-100 for the first time, in 0.4s. Compared with the prior best non-approximated solution (Usenix Security 2018), GForce speeds up non-linear~layers~in VGG by >34×. Our techniques shed light on a new direction that utilizes GPU throughout the model to minimize latency.

Modern processors allow attackers to leak data during transient (i.e., mis-speculated) execution through microarchitectural covert timing channels. While initial defenses were channel-specific, recent solutions employ speculative information flow control in an attempt to automatically mitigate attacks via any channel. However, we demonstrate that the current state-of-the-art defense fails to mitigate attacks using speculative stores, still allowing arbitrary data leakage during transient execution. Furthermore, we show that the state of the art does not scale to protect data in registers, incurring 30.8–63.4% overhead on SPEC 2017, depending on the threat model.

We then present DOLMA, the first defense to automatically provide comprehensive protection against all known transient execution attacks. DOLMA combines a lightweight speculative information flow control scheme with a set of secure performance optimizations. By enforcing a novel principle of transient non-observability, DOLMA ensures that a time slice on a core provides a unit of isolation in the context of existing attacks. Accordingly, DOLMA can allow speculative TLB/L1 cache accesses and variable-time arithmetic without loss of security. On SPEC 2017, DOLMA achieves comprehensive protection of data in memory at 10.2–29.7% overhead, adding protection for data in registers at 22.6–42.2% overhead (8.2–21.2% less than the state of the art, with greater security).

This paper presents the first empirical study of ground-truth data from a major underground shop selling stolen credit and debit cards. To date, there is little quantitative knowledge about how this segment of the underground economy operates, despite causing fraud losses estimated at billions of dollars a year. Our analysis of four years of leaked transactional data allows us to characterize this shop's business model, sellers, customers, and finances. The shop earned close to $104M in gross revenue and listed over 19M unique card numbers for sale. Around 97% of the inventory was stolen magnetic stripe data, commonly used to produce counterfeit cards for in-person payments. Perhaps surprisingly, customers purchased only 40% of this inventory. In contrast, the shop sold 83% of its card-not-present inventory, used for online fraud, which appeared to be in short supply. Demand and pricing were not uniform, as buyers appeared to perceive some banks as having weaker countermeasures against fraud. Even multiple years into the U.S. EMV chip deployment, the supply of stolen magnetic stripe data continued to increase sharply. In particular, we identified a continuing supply of newly issued cards not equipped with EMV chips, especially among prepaid cards. Our findings suggest that improvements to EMV chip deployment in the U.S., combined with a limited supply of stolen card-not-present data, could be avenues to decreasing the revenue and profitability of this shop.

Concession Abuse as a Service (CAaaS) is a growing scam service in underground forums that defrauds online retailers through the systematic abuse of their return policies (via social engineering) and the exploitation of loopholes in company protocols. Timely detection of such scams is difficult as they are fueled by an extensive suite of criminal services, such as credential theft, document forgery, and fake shipments. Ultimately, the scam enables malicious actors to steal arbitrary goods from merchants with minimal investment.

In this paper, we perform in-depth manual and automated analysis of public and private messages from four large underground forums to identify the malicious actors involved in CAaaS, carefully study the operation of the scam, and define attributes to fingerprint the scam and inform mitigation strategies. Additionally, we surveyed users to evaluate their attitudes toward these mitigations and understand the factors that merchants should consider before implementing these strategies. We find that the scam is easy to scale—and can bypass traditional anti-fraud efforts—and thus poses a notable threat to online retailers.