Technical Sessions

Privacy policies on websites are based on the notice-and-choice principle. They notify Web users of their privacy choices. However, many users do not read privacy policies or have difficulties understanding them. In order to increase privacy transparency we propose Privee—a software architecture for analyzing essential policy terms based on crowdsourcing and automatic classification techniques. We implement Privee in a proof of concept browser extension that retrieves policy analysis results from an online privacy policy repository or, if no such results are available, performs automatic classifications. While our classifiers achieve an overall F-1 score of 90%, our experimental results suggest that classifier performance is inherently limited as it correlates to the same variable to which human interpretations correlate—the ambiguity of natural language. This finding might be interpreted to call the notice-and-choice principle into question altogether. However, as our results further suggest that policy ambiguity decreases over time, we believe that the principle is workable. Consequently, we see Privee as a promising avenue for facilitating the notice-and-choice principle by accurately notifying Web users of privacy practices and increasing privacy transparency on the Web.

Users are increasingly storing, accessing, and exchanging data through public cloud services such as those provided by Google, Facebook, Apple, and Microsoft. Although users may want to have faith in cloud providers to provide good security protection, the confidentiality of any data in public clouds can be violated, and consequently, while providers may not be “doing evil,” we can not and should not trust them with data confidentiality.

To better protect the privacy of user data stored in the cloud, in this paper we propose a privacy-preserving system called Mimesis Aegis (M-Aegis) that is suitable for mobile platforms. M-Aegis is a new approach to user data privacy that not only provides isolation but also preserves the user experience through the creation of a conceptual layer called Layer 7.5 (L-7.5), which is interposed between the application (OSI Layer 7) and the user (Layer 8). This approach allows M-Aegis to implement true end-to-end encryption of user data with three goals in mind: 1) complete data and logic isolation from untrusted entities; 2) the preservation of original user experience with target apps; and 3) applicable to a large number of apps and resilient to app updates.

In order to preserve the exact application workflow and look-and-feel, M-Aegis uses L-7.5 to put a transparent window on top of existing application GUIs to both intercept plaintext user input before transforming the input and feeding it to the underlying app, and to reversetransform the output data from the app before displaying the plaintext data to the user. This technique allows MAegis to transparently integrate with most cloud services without hindering usability and without the need for reverse engineering. We implemented a prototype of MAegis on Android and show that it can support a number of popular cloud services, e.g. Gmail, Facebook Messenger, WhatsApp, etc.

Our performance evaluation and user study show that users incur minimal overhead when adopting M-Aegis on Android: imperceptible encryption/decryption latency and a low and adjustable false positive rate when searching over encrypted data.

Tor’s growing popularity and user diversity has resulted in network performance problems that are not well understood. A large body of work has attempted to solve these problems without a complete understanding of where congestion occurs in Tor. In this paper, we first study congestion in Tor at individual relays as well as along the entire end-to-end Tor path and find that congestion occurs almost exclusively in egress kernel socket buffers. We then analyze Tor’s socket interactions and discover two major issues affecting congestion: Tor writes sockets sequentially, and Tor writes as much as possible to each socket. We thus design, implement, and test KIST: a new socket management algorithm that uses real-time kernel information to dynamically compute the amount to write to each socket while considering all writable circuits when scheduling new cells. We find that, in the medians, KIST reduces circuit congestion by over 30 percent, reduces network latency by 18 percent, and increases network throughput by nearly 10 percent. We analyze the security of KIST and find an acceptable performance and security trade-off, as it does not significantly affect the outcome of well-known latency and throughput attacks. While our focus is Tor, our techniques and observations should help analyze and improve overlay and application performance, both for security applications and in general.

In response to increasingly sophisticated state-sponsored Internet censorship, recent work has proposed a new approach to censorship resistance: end-to-middle proxying. This concept, developed in systems such as Telex, Decoy Routing, and Cirripede, moves anticensorship technology into the core of the network, at large ISPs outside the censoring country. In this paper, we focus on two technical obstacles to the deployment of certain end-to-middle schemes: the need to selectively block flows and the need to observe both directions of a connection. We propose a new construction, TapDance, that removes these requirements. TapDance employs a novel TCP-level technique that allows the anticensorship station at an ISP to function as a passive network tap, without an inline blocking component. We also apply a novel steganographic encoding to embed control messages in TLS ciphertext, allowing us to operate on HTTPS connections even under asymmetric routing. We implement and evaluate a TapDance prototype that demonstrates how the system could function with minimal impact on an ISP’s network operations.

State-of-the-art memory forensics involves signature-based scanning of memory images to uncover data structure instances of interest to investigators. A largely unaddressed challenge is that investigators may not be able to interpret the content of data structure fields, even with a deep understanding of the data structure’s syntax and semantics. This is very common for data structures with application-specific encoding, such as those representing images, figures, passwords, and formatted file contents. For example, an investigator may know that a buffer field is holding a photo image, but still cannot display (and hence understand) the image. We call this the data structure content reverse engineering challenge. In this paper, we present DSCRETE, a system that enables automatic interpretation and rendering of in-memory data structure contents. DSCRETE is based on the observation that the application in which a data structure is defined usually contains interpretation and rendering logic to generate human-understandable output for that data structure. Hence DSCRETE aims to identify and reuse such logic in the program’s binary and create a “scanner+renderer” tool for scanning and rendering instances of the data structure in a memory image. Different from signature-based approaches, DSCRETE avoids reverse engineering data structure signatures. Our evaluation with a wide range of real-world application binaries shows that DSCRETE is able to recover a variety of application data—e.g., images, figures, screenshots, user accounts, and formatted files and messages—with high accuracy. The raw contents of such data would otherwise be unfathomable to human investigators.

The volume and the sophistication of malware are continuously increasing and evolving. Automated dynamic malware analysis is a widely-adopted approach for detecting malicious software. However, many recent malware samples try to evade detection by identifying the presence of the analysis environment itself, and refraining from performing malicious actions. Because of the sophistication of the techniques used by the malware authors, so far the analysis and detection of evasive malware has been largely a manual process. One approach to automatic detection of these evasive malware samples is to execute the same sample in multiple analysis environments, and then compare its behaviors, in the assumption that a deviation in the behavior is evidence of an attempt to evade one or more analysis systems. For this reason, it is important to provide a reference system (often called bare-metal) in which the malware is analyzed without the use of any detectable component.

In this paper, we present BareCloud, an automated evasive malware detection system based on bare-metal dynamic malware analysis. Our bare-metal analysis system does not introduce any in-guest monitoring component into the malware execution platform. This makes our approach more transparent and robust against sophisticated evasion techniques. We compare the malware behavior observed in the bare-metal system with other popular malware analysis systems. We introduce a novel approach of hierarchical similarity-based malware behavior comparison to analyze the behavior of a sample in the various analysis systems. Our experiments show that our approach produces better evasion detection results compared to previous methods. BareCloud was able to automatically detect 5,835 evasive malware out of 110,005 recent samples.

This paper analyzes the actual cost of attacking TLS implementations that use NIST’s Dual EC pseudorandom number generator, assuming that the attacker generated the constants used in Dual EC. It has been known for several years that an attacker generating these constants and seeing a long enough stretch of Dual EC output bits can predict all future outputs; but TLS does not naturally provide a long enough stretch of output bits, and the cost of an attack turns out to depend heavily on choices made in implementing the RNG and on choices made in implementing other parts of TLS.

Specifically, this paper investigates OpenSSL-FIPS, Windows’ SChannel, and the C/C++ and Java versions of the RSA BSAFE library. This paper shows that Dual EC exploitability is fragile, and in particular is stopped by an outright bug in the certified Dual EC implementation in OpenSSL. On the other hand, this paper also shows that Dual EC exploitability benefits from a modification made to the Dual EC standard in 2007; from several attack optimizations introduced here; and from various proposed TLS extensions, one of which is implemented in BSAFE, though disabled in the version we obtained and studied. The paper’s attacks are implemented; benchmarked; tested against libraries modified to use new Dual EC constants; and verified to successfully recover TLS plaintext.

In the attempt to bring modern broadband Internet features to traditional broadcast television, the Digital Video Broadcasting (DVB) consortium introduced a specification called Hybrid Broadcast-Broadband Television (HbbTV), which allows broadcast streams to include embedded HTML content which is rendered by the television. This system is already in very wide deployment in Europe, and has recently been adopted as part of the American digital television standard.

Our analyses of the specifications, and of real systems implementing them, show that the broadband and broadcast systems are combined insecurely. This enables a large-scale exploitation technique with a localized geographical footprint based on radio frequency (RF) injection, which requires a minimal budget and infrastructure and is remarkably difficult to detect. Despite our responsible disclosure to the standards body, our attack was viewed as too expensive and with limited pay-off to the attackers.

In this paper, we present the attack methodology and a number of follow-on exploitation techniques that provide significant flexibility to attackers. Furthermore, we demonstrate that the technical complexity and required budget are low, making this attack practical and realistic, especially in areas with high population density – in a dense urban area, an attacker with a budget of about $450 can target more than 20,000 devices in a single attack. A unique aspect of this attack is that, in contrast to most Internet of Things/Cyber-Physical System threat scenarios where the attack comes from the data network side and affects the physical world, our attack uses the physical broadcast network to attack the data network.

We study the security of popular password managers and their policies on automatically filling in Web passwords. We examine browser built-in password managers, mobile password managers, and 3rd party managers. We observe significant differences in autofill policies among password managers. Several autofill policies can lead to disastrous consequences where a remote network attacker can extract multiple passwords from the user’s password manager without any interaction with the user. We experiment with these attacks and with techniques to enhance the security of password managers. We show that our enhancements can be adopted by existing managers.

This paper presents SpanDex, a set of extensions to Android’s Dalvik virtual machine that ensures apps do not leak users’ passwords. The primary technical challenge addressed by SpanDex is precise, sound, and efficient handling of implicit information flows (e.g., information transferred by a program’s control flow). SpanDex handles implicit flows by borrowing techniques from symbolic execution to precisely quantify the amount of information a process’ control flow reveals about a secret. To apply these techniques at runtime without sacrificing performance, SpanDex runs untrusted code in a data-flow sensitive sandbox, which limits the mix of operations that an app can perform on sensitive data. Experiments with a SpanDex prototype using 50 popular Android apps and an analysis of a large list of leaked passwords predicts that for 90% of users, an attacker would need over 80 login attempts to guess their password. Today the same attacker would need only one attempt for all users.

Users speaking different languages may prefer different patterns in creating their passwords, and thus knowledge on English passwords cannot help to guess passwords from other languages well. Research has already shown Chinese passwords are one of the most difficult ones to guess. We believe that the conclusion is biased because, to the best of our knowledge, little empirical study has examined regional differences of passwords on a large scale, especially on Chinese passwords. In this paper, we study the differences between passwords from Chinese and English speaking users, leveraging over 100 million leaked and publicly available passwords from Chinese and international websites in recent years. We found that Chinese prefer digits when composing their passwords while English users prefer letters, especially lowercase letters. However, their strength against password guessing is similar. Second, we observe that both users prefer to use the patterns that they are familiar with, e.g., Chinese Pinyins for Chinese and English words for English users. Third, we observe that both Chinese and English users prefer their conventional format when they use dates to construct passwords. Based on these observations, we improve a PCFG (Probabilistic Context-Free Grammar) based password guessing method by inserting Pinyins (about 2.3% more entries) into the attack dictionary and insert our observed composition rules into the guessing rule set. As a result, our experiments show that the efficiency of password guessing increases by 34%.

To discourage the creation of predictable passwords, vulnerable to guessing attacks, we present Telepathwords. As a user creates a password, Telepathwords makes realtime predictions for the next character that user will type. While the concept is simple, making accurate predictions requires efficient algorithms to model users’ behavior and to employ already-typed characters to predict subsequent ones. We first made the Telepathwords technology available to the public in late 2013 and have since served hundreds of thousands of user sessions.

We ran a human-subjects experiment to compare password policies that use Telepathwords to those that rely on composition rules, comparing participants’ passwords using two different password-evaluation algorithms. We found that participants create far fewer weak passwords using the Telepathwords-based policies than policies based only on character composition. Participants using Telepathwords were also more likely to report that the password feedback was helpful.

Public infrastructure-as-a-service clouds, such as Amazon EC2 and Microsoft Azure allow arbitrary clients to run virtual machines (VMs) on shared physical infrastructure. This practice of multi-tenancy brings economies of scale, but also introduces the threat of malicious VMs abusing the scheduling of shared resources. Recent works have shown how to mount cross- VM side-channel attacks to steal cryptographic secrets. The straightforward solution is hard isolation that dedicates hardware to each VM. However, this comes at the cost of reduced efficiency.

We investigate the principle of soft isolation: reduce the risk of sharing through better scheduling. With experimental measurements, we show that a minimum run time (MRT) guarantee for VM virtual CPUs that limits the frequency of preemptions can effectively prevent existing Prime+Probe cache-based side-channel attacks. Through experimental measurements, we find that the performance impact of MRT guarantees can be very low, particularly in multi-core settings. Finally, we integrate a simple per-core CPU state cleansing mechanism, a form of hard isolation, into Xen. It provides further protection against side-channel attacks at little cost when used in conjunction with an MRT guarantee.

Sharing memory pages between non-trusting processes is a common method of reducing the memory footprint of multi-tenanted systems. In this paper we demonstrate that, due to a weakness in the Intel X86 processors, page sharing exposes processes to information leaks. We present FLUSH+RELOAD, a cache side-channel attack technique that exploits this weakness to monitor access to memory lines in shared pages. Unlike previous cache side-channel attacks, FLUSH+RELOAD targets the Last- Level Cache (i.e. L3 on processors with three cache levels). Consequently, the attack program and the victim do not need to share the execution core.

We demonstrate the efficacy of the FLUSH+RELOAD attack by using it to extract the private encryption keys from a victim program running GnuPG 1.4.13. We tested the attack both between two unrelated processes in a single operating system and between processes running in separate virtual machines. On average, the attack is able to recover 96.7% of the bits of the secret key by observing a single signature or decryption round.

We present Burst ORAM, the first oblivious cloud storage system to achieve both practical response times and low total bandwidth consumption for bursty workloads. For real-world workloads, Burst ORAM can attain response times that are nearly optimal and orders of magnitude lower than the best existing ORAM systems by reducing online bandwidth costs and aggressively rescheduling shuffling work to delay the bulk of the IO until idle periods.

We evaluate our design on an enterprise file system trace with about 7,500 clients over a 15 day period, comparing to an insecure baseline encrypted block store without ORAM. We show that when baseline response times are low, Burst ORAM response times are comparably low. In a 32TB ORAM with 50ms network latency and sufficient bandwidth capacity to ensure 90% of requests have baseline response times under 53ms, 90% of Burst ORAM requests have response times under 63ms, while requiring only 30 times the total bandwidth consumption of the insecure baseline. Similarly, with sufficient bandwidth to ensure 99:9% of requests have baseline responses under 70ms, 99:9% of Burst ORAM requests have response times under 76ms.

We build a system that provides succinct non-interactive zero-knowledge proofs (zk-SNARKs) for program executions on a von Neumann RISC architecture. The system has two components: a cryptographic proof system for verifying satisfiability of arithmetic circuits, and a circuit generator to translate program executions to such circuits. Our design of both components improves in functionality and efficiency over prior work, as follows.

Our circuit generator is the first to be universal: it does not need to know the program, but only a bound on its running time. Moreover, the size of the output circuit depends additively (rather than multiplicatively) on program size, allowing verification of larger programs.

The cryptographic proof system improves proving and verification times, by leveraging new algorithms and a pairing library tailored to the protocol.

We evaluated our system for programs with up to 10;000 instructions, running for up to 32;000 machine steps, each of which can arbitrarily access random-access memory; and also demonstrated it executing programs that use just-in-time compilation. Our proofs are 230 bytes long at 80 bits of security, or 288 bytes long at 128 bits of security. Typical verification time is 5 ms, regardless of the original program’s running time.

Encryption schemes where the ciphertext must abide by a specified format have diverse applications, ranging from in-place encryption in databases to per-message encryption of network traffic for censorship circumvention. Despite this, a unifying framework for deploying such encryption schemes has not been developed. One consequence of this is that current schemes are ad-hoc; another is a requirement for expert knowledge that can disuade one from using encryption at all.

We present a general-purpose library (called libfte) that aids engineers in the development and deployment of format-preserving encryption (FPE) and formattransforming encryption (FTE) schemes. It incorporates a new algorithmic approach for performing FPE/FTE using the nondeterministic finite-state automata (NFA) representation of a regular expression when specifying formats. This approach was previously considered unworkable, and our approach closes this open problem. We evaluate libfte and show that, compared to other encryption solutions, it introduces negligible latency overhead, and can decrease diskspace usage by as much as 62.5% when used for simultaneous encryption and compression in a PostgreSQL database (both relative to conventional encryption mechanisms). In the censorship circumvention setting we show that, using regularexpression formats lifted from the Snort IDS, libfte can reduce client/server memory requirements by as much as 30%.

Traditionally, confidentiality and integrity have been two desirable design goals that are have been dicult to combine. Zero-Knowledge Proofs of Knowledge (ZKPK) offer a rigorous set of cryptographic mechanisms to balance these concerns. However, published uses of ZKPK have been dicult for regular developers to integrate into their code and, on top of that, have not been demonstrated to scale as required by most realistic applications.

This paper presents ZØ (pronounced “zee-not”), a compiler that consumes applications written in C# into code that automatically produces scalable zeroknowledge proofs of knowledge, while automatically splitting applications into distributed multi-tier code. ZØ builds detailed cost models and uses two existing zeroknowledge back-ends with varying performance characteristics to select the most ecient translation. Our case studies have been directly inspired by existing sophisticated widely-deployed commercial products that require both privacy and integrity. The performance delivered by ZØ is as much as 40 faster across six complex applications. We find that when applications are scaled to real-world settings, existing zero-knowledge compilers often produce code that fails to run or even compile in a reasonable amount of time. In these cases, ZØ is the only solution we know about that is able to provide an application that works at scale.

Android, iOS, and Windows 8 are changing the application architecture of consumer operating systems. These new architectures required OS designers to rethink security and access control. While the new security architectures improve on traditional desktop and server OS designs, they lack sufficient protection semantics for different classes of OS customers (e.g., consumer, enterprise, and government). The Android OS in particular has seen over a dozen research proposals for security enhancements. This paper seeks to promote OS security extensibility in the Android OS. We propose the Android Security Modules (ASM) framework, which provides a programmable interface for defining new reference monitors for Android. We drive the ASM design by studying the authorization hook requirements of recent security enhancement proposals and identify that new OSes such as Android require new types of authorization hooks (e.g., replacing data). We describe the design and implementation of ASM and demonstrate its utility by developing reference monitors called ASM apps. Finally, ASM is not only beneficial for security researchers. If adopted by Google, we envision ASM enabling in-thefield security enhancement of Android devices without requiring root access, a significant limitation of existing bring-your-own-device solutions.

The security of smartphone GUI frameworks remains an important yet under-scrutinized topic. In this paper, we report that on the Android system (and likely other OSes), a weaker form of GUI confidentiality can be breached in the form of UI state (not the pixels) by a background app without requiring any permissions. Our finding leads to a class of attacks which we name UI state inference attack. The underlying problem is that popular GUI frameworks by design can potentially reveal every UI state change through a newly-discovered public side channel — shared memory. In our evaluation, we show that for 6 out of 7 popular Android apps, the UI state inference accuracies are 80–90% for the first candidate UI states, and over 93% for the top 3 candidates.

Even though the UI state does not reveal the exact pixels, we show that it can serve as a powerful building block to enable more serious attacks. To demonstrate this, we design and fully implement several new attacks based on the UI state inference attack, including hijacking the UI state to steal sensitive user input (e.g., login credentials) and obtain sensitive camera images shot by the user (e.g., personal check photos for banking apps). We also discuss non-trivial challenges in eliminating the identified side channel, and suggest more secure alternative system designs.