TL;DR
This paper introduces a novel multi-agent system for network attack auditing that improves upon traditional centralized approaches. By distributing audit tasks among specialized agents (network attack, host, and management), the system aims to enhance efficiency, precision, and real-time detection of various network threats.
What Problem Does This Paper Address?
The paper tackles the growing challenge of network security, particularly internal threats which, as of 1999 CSI data, accounted for 82% of network security issues. Existing security measures like firewalls and intrusion detection systems are often supplemented by network safety auditing systems. However, traditional centralized audit structures can suffer from reduced efficiency and precision due to the sheer volume of data. The authors propose a distributed, multi-agent approach to overcome these limitations, aiming for more effective real-time and historical data analysis, irregularity detection, and warning capabilities within local area networks (LANs).
Methods
The core of the proposed solution is a multi-agent system architecture comprising three types of agents:
- Network Attack Audit Agent (AA): Deployed near servers, this agent monitors network packets in promiscuous mode. It utilizes an improved information entropy detection algorithm to identify abrupt changes in packet header entropy, signaling potential attacks.
- Host Audit Agent (HA): Distributed across individual hosts, this agent analyzes system logs and network packet contact information. It employs association rule mining and a rule library to identify abnormal host operating behavior or network access patterns.
- Management Control Center Agent (MCA): This central agent aggregates alarms and audit information from AA and HA, overseeing the entire LAN's audit work. It manages and analyzes audit results, providing timely insights and facilitating preventive measures.
A key innovation is the improved information entropy detection algorithm. While traditional entropy-based methods for DDoS detection exist, this paper refines the approach by considering multiple packet header fields (ITEMS: Time, SrcIP, DstIP, SrcPort, DstPort, ProtocolType, TTL, ACK, SYN, TcpWin, IcmpCode) and their combinations. This allows for a more nuanced analysis and detection of abnormal events, reducing false alarms and improving accuracy compared to methods relying solely on mathematical statistics or single-field entropy. The algorithm also aims to reduce computational complexity by optimizing the update process of entropy values within a sliding window.
Datasets
Not discussed in the paper. The authors mention using Tamir.IPLib.SharpPcap for network packet collection and implementing the system on Windows Server 2003 and Windows XP operating systems.
Key Insights and Ideas
- Distributed Auditing is Key: Centralized systems struggle with the sheer volume of network data. A distributed, multi-agent approach allows for specialized, localized analysis, significantly improving efficiency and reducing server load.
- Information Entropy as a Sensitive Indicator: Abrupt changes in information entropy within network packet headers are highly sensitive indicators of malicious activity, particularly for attacks like DDoS.
- Multi-Field Entropy Analysis Enhances Detection: Moving beyond single-field entropy analysis to consider combinations of packet header fields (ITEMS) provides a richer context for detecting anomalies and reduces the likelihood of false positives.
- Agent Specialization for Targeted Auditing: Assigning specific roles to agents (network monitoring, host behavior analysis, central management) streamlines the audit process and allows for more focused and effective threat detection.
Key Results
The paper demonstrates the system's effectiveness in detecting various attacks, including UDPFlood, ACK Flood, SynFlood, LAND, and Ping of Death. Specifically, Figure 3 shows a sharp fall in entropy values for DstIP, DstPort, and ProtocolType during a UDPFlood attack, pinpointing the attack's characteristics (UDP, 192.168.0.7, port 80). Table 1 quantifies these entropy changes, showing significant drops in values between time seconds 19 and 20. The authors claim the system can effectively detect and prevent IP and ARP spoofing as well.
Limitations and Future Work
- Computational Complexity: While the improved entropy algorithm aims to reduce complexity, the paper acknowledges that recalculating probabilities and entropies can still increase computational load, especially with larger window sizes.
- Real-time Performance: The paper focuses on the audit system's design and detection capabilities. Further work could explore optimizing real-time performance and scalability for very large networks.
- Specific Attack Signatures: While the entropy-based approach is general, further research could focus on developing more specific detection rules and signatures for a wider range of sophisticated attacks.
- Agent Communication and Coordination: The paper outlines the roles of agents but could benefit from a deeper dive into the communication protocols and coordination mechanisms between agents for more complex scenarios.
Summary updated