Strategic Alert Throttling for Intrusion Detection Systems

Gianni Tedesco, Uwe AickelinSsrn 2005 · DOI · OpenReview

TL;DR

Alert flood attacks overwhelm Intrusion Detection Systems (IDS) by flooding them with false alerts, masking real threats. This paper proposes a novel approach that combines real-time alert correlation with a token bucket filter to strategically throttle excess alerts, significantly improving IDS capacity and resilience without missing critical security events.

What Problem Does This Paper Address?

Modern networks rely heavily on Intrusion Detection Systems (IDS) to monitor for malicious activity. However, these systems themselves are becoming targets. "Alert flood attacks" exploit a critical vulnerability: the communication channel used to report alerts. Attackers can inundate an IDS with a deluge of false alerts, effectively overwhelming its capacity to process and report genuine threats. This makes it incredibly difficult for security analysts to distinguish real attacks from the noise, allowing attackers to hide their malicious activities. While alert correlation techniques exist to reduce false positives, they can also be overwhelmed by a flood of alerts. The challenge is to maintain the IDS's ability to detect and report real threats even under a sustained alert flood attack, without simply discarding potentially important information.

Methods

The paper introduces a novel algorithm that combines two key techniques:

  • Token Bucket Filter (TBF): This is a rate-limiting algorithm. It works by dispensing "tokens" at a fixed rate. Alerts can only pass through the filter if a token is available. This effectively caps the rate of alerts being processed. The "burstiness" feature of TBFs allows for temporary spikes in alert rates, but sustained high rates will lead to alerts being dropped.
  • Attack Graph Correlation: This technique uses a model of possible attacker strategies (an attack graph) to analyze and correlate alerts. The key innovation here is using the attack graph to inform the throttling process. Instead of blindly throttling all alerts, the system prioritizes alerts that are crucial for building a complete attack scenario within the attack graph. This ensures that strategically important alerts are less likely to be dropped.

The proposed system places the token bucket filter after the signature matching but before the alerting component, with the correlation algorithm acting as an intermediary. This allows the correlation algorithm to decide which alerts are important enough to pass through the throttled output. For alerts that don't map to specific exploit vertices in the attack graph, a separate token bucket filter is applied on a per-signature basis.

Datasets

Not discussed in the paper. The experiments used a tcpdump capture file of an ICMP flood attack from the "Firestorm" IDS system.

Key Insights and Ideas

  • Strategic Throttling is Key: Simply increasing IDS capacity isn't always feasible, especially for human operators. Throttling is necessary, but it must be intelligent.
  • Attack Graphs Provide Context for Throttling: The attack graph is not just for correlation; it provides the crucial context to decide which alerts are important and should not be dropped during a flood. This prevents the throttling mechanism from inadvertently helping attackers by discarding vital evidence.
  • Minimizing Resource Exhaustion: The core idea is to reduce the resources required by the IDS to handle an alert flood, making the attack prohibitively expensive for the attacker.
  • Handling Unmapped Alerts: The paper acknowledges that not all alerts will map to the attack graph and proposes a per-signature throttling mechanism for these cases, highlighting a trade-off between data fidelity and efficiency.
  • Run-Length Encoding (RLE) for Dropped Alerts: To provide some visibility into dropped alerts, the paper suggests using RLE to represent sequences of identical, throttled alerts, allowing for minimal reconstruction of lost information.

Key Results

The paper presents experimental results using a simulated alert flood attack (ICMP flood at ~7,343 packets/sec) against the "Firestorm" IDS.

  • Data Logging Reduction: The amount of data logged was reduced by "several orders of magnitude."
  • Runtime Improvement: The run time decreased by approximately 30%, while CPU time only decreased by about 10%. This suggests that the IDS spends less time waiting for I/O completion when throttling is enabled.
  • Alert Output Reduction: The number of alerts output was reduced by orders of magnitude.

The authors note that these results are particularly impactful in real-world deployments where alerts are transmitted over a network, adding further latency and bandwidth constraints.

Limitations and Future Work

  • Optimal TBF Configuration: The paper acknowledges that the chosen parameters for the token bucket filters (2 alerts/sec, burst of 20) were arbitrary and suggests that further investigation is required to determine optimal configurations.
  • Handling Unmapped Alerts: The paper notes that a more prudent approach was taken for alerts not mapping to exploit vertices, but further work is needed on how best to handle these.
  • Data Fidelity vs. Efficiency Trade-off: The authors recognize that there's an inherent trade-off between keeping all alert data and improving efficiency, and further research is needed to balance this.
  • Real-world Deployment: While the experiments show promising results, the paper implies that real-world deployment might present additional challenges and opportunities for performance gains.

Summary updated

Abstract

Network intrusion detection systems are themselves becoming targets of attackers. Alert flood attacks may be used to conceal malicious activity by hiding it among a deluge of false alerts sent by the attacker. Although these types of attacks are very hard to stop completely, our aim is to present techniques that improve alert throughput and capacity to such an extent that the resources required to successfully mount the attack become prohibitive. The key idea presented is to combine a token bucket filter with a realtime correlation algorithm. The proposed algorithm throttles alert output from the IDS when an attack is detected. The attack graph used in the correlation algorithm is used to make sure that alerts crucial to forming strategies are not discarded by throttling.
BibTeX
@misc{tedesco2005strategic,
  author = {Gianni Tedesco and Uwe Aickelin},
  title = {{Strategic Alert Throttling for Intrusion Detection Systems}},
  year = {2005},
  eprint = {0801.4119},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/0801.4119},
  howpublished = {Ssrn},
}
Figure