Session State Transition Based Large Network IDS

In order to present large-scale malicious attacks on an ISP network to maintain network services, we have designed a method to record key packets classified by sessions. Session is the service provided above the IP layer. We define a TCP connection a session, a UDP packet exchange a session, or echo and echo response of ICMP to be a session. The research of network attack/intrusion/information collection has shown that most of the illegal action performed would have something special ongoing in such sessions. For example, winnuke will send OOB packets to the 139 port of a host; most of the platform detection will use strange packets too. Not only the strange packets itself, but the sequence of such packets going through the network indicate the attack. For example, teardrop will transmit packets that have abnormal fragment offset in the second packet, then cause some platform to crash. Some patterns of sessions will be created by flood based attack/information collection. For example, the SYN flood will create a pile SYN-SYN ACK-RST packets in the network, and most of scan tools will create several kind of patterns in the network, all of these patterns indicate the failure of the connection, these include SYN-SYN ACK-RST and SYN-RST and SYNICMP Unreachable message. Based on this thought, we have designed the session-state transition analysis. We will define some packets as the indication of the session state. The happening of such packets causes the change of the session state. When comparing with the predefined rules, we will detect most of the DOS attacks. Another approach is to store these session states transition patterns into a database; thus we can calculate the happening rates of some specific patterns. Compared with the average level, abnormal high happening rates often indicate the possible attack or information collection. For example, we can collect a site’s all sessions' SYN-SYN ACK-RST pattern to decide whether a normal scan had happened. The implementation includes four parts. The first is the data collection part, which collects and unwraps packets passing through the network; the second part is the signature matching part, which will match the packet signature, to filter only the specified packets; the third part will cluster such pa ckets into sessions, and store the session specific signature chain and check whether a rule based match is satisfied; the fourth part will flush the session data into a database, and check whether a statistical based anomaly has happened. Using such kind of techniques has several basic advantages. The first is not to violate privacy, since we are interested in only packet header to know whether a state has changed, to inspect header only also make this implementation efficient and fit for a large scale network. The other advantage is to avoid the headache to set the threshold of a statistical approach. Most scan detection tools (For example, gabriel) will calculate the burst of connections. New scan technique has appeared to avoid burst of connections, for example, slow scan and stealthy scan. Set a proper threshold is much more difficult for a large-scale network. For rule based analysis, since we use the state transition to detect intrusion, we could predict the happening of some attacks in a premature stage. The future approach includes the content analysis based IDS, especially the remote buffer overflow detection. This part of research is underway.