Synthesizing Fast Intrusion Prevention/Detection Systems from High-Level Specifications

To build survivable information systems (i.e., systems that continue to provide their services in spite of coordinated attacks), it is necessary to detect and isolate intrusions before they impact system performance or functionality. Previous research in this area has focussed primarily on detecting intrusions after the fact, rather than preventing them in the first place. We have developed a new approach based on specifying intended program behaviors using patterns over sequences of system calls. The patterns can also capture conditions on the values of system-call arguments. At runtime, we intercept the system calls made by processes, compare them against specifications, and disallow (or otherwise modify) those calls that deviate from specifications. Since our approach is capable of modifying a system call before it is delivered to the operating system kernel, it is capable of reacting before any damage-causing system call is executed by a process under attack. We present our specification language and illustrate its use by developing a specification for the ftp server. Observe that in our approach, every system call is intercepted and subject to potentially expensive operations for matching against many patterns that specify normal/abnormal behavior. Thus, minimizing the overheads incurred for pattern-matching is critical for the viability of our approach. We solve this problem by developing a new, low-overhead algorithm for matching runtime behaviors against specifications. A salient feature of our algorithm is that its runtime is almost independent of the number of patterns. In most cases, it uses a constant amount of time per system call intercepted, and uses a constant amount of storage, both independent of either the size or number of patterns. These benefits make our algorithm useful for many other intrusion detection methods that employ pattern-matching. We describe our algorithm, and evaluate its performance through experiments.