Abstract: The Dendritic Cell Algorithm (DCA) is an immune-inspired algorithm developed for anomaly detection through multi-sensor data fusion and correlation. Previous applications have shown success in detecting malicious port scanning. This work compares the performance of the DCA against a Self-Organizing Map (SOM) in detecting SYN port scans. Results indicate that the DCA performs at a standard comparable to established techniques like the SOM, even showing statistical advantages in specific high-noise scenarios.
Introduction
Artificial Immune Systems (AIS) aim to replicate the biological immune system's robust detection capabilities. Early AIS models often struggled with scalability and high false-positive rates. The Dendritic Cell Algorithm (DCA) addresses these issues by incorporating "Danger Theory," focusing on context-aware detection. This study benchmarks the DCA against the Self-Organizing Map (SOM), a well-regarded neural-inspired clustering method, to evaluate its efficacy in a cybersecurity context.
Methodology
The comparison focuses on two distinct biologically-inspired paradigms:
- Dendritic Cell Algorithm (DCA): Inspired by innate immunity, it fuses three signal types—PAMPs (certain signs of anomaly), Danger Signals (probable signs of stress), and Safe Signals (indicators of normal behavior)—to calculate a Mature Context Antigen Value (MCAV) for specific processes (antigens).
- Self-Organizing Map (SOM): An unsupervised learning technique that maps high-dimensional data to a 2D grid, identifying anomalies based on the distance to the Best Matching Unit (BMU).
To ensure parity, data was segmented into fixed-size windows ($z$), and the Mann-Whitney test was employed for statistical validation of the performance differences.
Experimental Setup and Datasets
The study utilized two primary datasets totaling approximately 13 million data items:
- Passive Normal (PN): A SYN port scan (using
nmap) against a system with minimal background activity. - Active Normal (AN): A SYN scan obscured by heavy background traffic, including web browsing, chat sessions, and email via SSH.
Signals were extracted from the Linux /proc filesystem and network statistics, focusing on metrics like ICMP error rates, TCP reset packets, and network packet rates.
Discussion and Key Insights
The DCA's primary strength is its "context-aware" correlation, linking signal spikes directly to specific process IDs (antigens). Rather than just observing a traffic spike, it evaluates if that spike is associated with a specific actor.
Interestingly, both algorithms exhibited shared blind spots, producing false positives for the same normal system processes. This suggests that some benign behaviors naturally mimic attack patterns regardless of the detection logic. However, the DCA demonstrated a tighter coupling between stress signals and offending processes, providing a statistically significant advantage in identifying the nmap process specifically during periods of high background noise.
Results
Both algorithms proved effective at identifying SYN-based port scans, significantly outperforming a baseline k-means classifier (which achieved only ~52% accuracy). In the "Active Normal" scenario, the DCA's detection of the nmap process was statistically superior to the SOM, yielding a p-value of 0.0001. This indicates that the DCA can produce anomaly detection results to the same standard as, or better than, established techniques.
Conclusion and Future Work
The DCA is a viable and competitive alternative to traditional machine learning methods for network anomaly detection. While it excels in context-aware scenarios like port scanning, future work should explore its performance on larger, more diverse datasets. Key areas for development include moving toward dynamic segmentation windows that adapt to system activity and establishing a deeper mathematical grounding for the DCA’s data fusion properties.
Summary updated