TL;DR
This paper introduces a robust methodology for detecting and tracing cyberattacks that use "stepping stones" – a chain of compromised intermediate computers – to hide the attacker's origin. By combining traditional stepping stone detection with novel anomaly detection techniques for jitter and chaff, the authors create a system that is significantly harder for attackers to evade, achieving a 99% success rate.
What Problem Does This Paper Address?
Cyberattacks often originate from compromised machines to mask the attacker's true identity. A common technique is the use of "stepping stones," where an attacker chains together multiple compromised machines to reach a target. Traditional methods for detecting these stepping stones often rely on timing correlations between network traffic flows. However, attackers can actively evade these methods by introducing "jitter" (small, random delays) or "chaff" (extra, irrelevant packets) into the traffic. This paper addresses the challenge of detecting stepping stones in the presence of such active evasive maneuvers, particularly from insider threats.
Methods
The core of this paper is the development of three anomaly detection algorithms designed to identify traffic patterns indicative of jitter and chaff:
- Response-time based algorithm: This method detects jitter by monitoring the time it takes for a server to respond to a client's request. If the response time exceeds a certain threshold (RTT + SRT), the connection is flagged as anomalous.
- Edit-distance based algorithm: This algorithm detects chaff by analyzing the sequences of ON and OFF periods in bidirectional network traffic. Normal connections exhibit similar sequences; injecting chaff makes these sequences dissimilar, leading to a positive edit distance that flags the connection as anomalous.
- Causality based algorithm: This method also targets chaff by examining the causal relationship between packet flows in both directions of a connection. In normal interactive traffic, there's a predictable pattern of ON periods (e.g., one response for each command). Deviations from this pattern (zero or more than one response) indicate anomalous activity, likely due to chaff.
These anomaly detection algorithms are then integrated with existing timing-based stepping stone detection methods to create a more resilient attacker traceback framework.
Datasets
The authors evaluated their algorithms using:
- Real-world Internet traces: Collected from Telcordia Technologies and the University of Texas Computer Sciences department.
- Custom-generated traces: Created on the Deter test bed, specifically designed to inject jitter and chaff into network connections.
- Live traffic: Tested on the Deter test bed.
The paper does not introduce a new benchmark dataset but rather uses existing real-world traces and a controlled experimental setup on Deter to validate their approach.
Key Insights and Ideas
- Active evasion is a significant threat: Attackers can easily bypass older stepping stone detection methods by simply introducing jitter or chaff.
- Anomaly detection is key to robustness: By looking for deviations from "normal" interactive traffic patterns (response times, sequence similarity, causality), the proposed algorithms can detect evasive maneuvers that timing-based methods miss.
- Combining techniques is powerful: The synergy between stepping stone detection and anomaly detection creates a framework that is much harder to evade than either approach alone. An attacker trying to hide from one method might inadvertently trigger the other.
- Insider threats are a prime target: The assumptions made by the algorithms (e.g., software deployment, node cooperation) are more likely to hold in intranet environments, making them particularly effective against insider attacks.
Key Results
The proposed integrated methodology achieved a high success rate of approximately 99%.
- Response-time based algorithm: Demonstrated very low false positives and negatives, with a success rate of 98.99%. It correctly identified jitter in controlled tests and had only a few false positives on real-world traces.
- Edit-distance based algorithm: Achieved a success rate of 98.81%. It effectively detected chaff in controlled tests but had a higher number of false positives on real-world traces.
- Causality based algorithm: Showed the highest success rate at 99.06%. It was effective in detecting chaff and had minimal false positives.
Limitations and Future Work
- Magic Numbers: The algorithms rely on "magic numbers" (parameters like
Tidle,δ,γ,β,εRT) that were determined through extensive testing. Dynamically determining these parameters could further improve performance. - Assumptions for Intranets: While effective for intranets, the assumptions about software deployment and node cooperation would require significant architectural changes for broader internet deployment.
- Aggressive Attackers: The authors acknowledge that a "sufficiently aggressive attacker" might still find ways to evade the system, though it significantly raises the bar.
- Port Monitoring: The current Bro implementation monitors specific ports. Future work could involve identifying anomalous activity on any port, not just those typically used for interactive traffic.
- Dynamic Parameter Tuning: Exploring methods for dynamically adjusting algorithm parameters based on observed traffic characteristics is a promising avenue.
Summary updated