High-performance network intrusion detection system

In this paper we present a new approach for network intrusion detection based on concise specifications that characterize normal and abnormal network packet sequences. Our specification language is geared for a robust network intrusion detection by enforcing a strict type discipline via a combination of static and dynamic type checking. Unlike most previous approaches in network intrusion detection, our approach can easily support new network protocols as information relating to the protocols are not hard-coded into the system. Instead, we simply add suitable type definitions in the specifications and define intrusion patterns on these types. We compile these specifications into a high-performance network intrusion detection system. Important components of our approach include efficient algorithms for pattern-matching and information aggregation on sequences of network packets. In particular, our techniques ensure that the matching time is insensitive to the number of patterns characterizing different network intrusions, and that the aggregation operations typically take constant time per packet. Our system participated in an intrusion detection evaluation organized by MIT Lincoln Labs, where our system demonstrated its effectiveness (96% detection rate on low-level network attacks) and performance (real-time detection at 500Mbps), while producing very few false positives (0.05 to 0.1 per attack).