Abstract
The hidden network environment of multi-agent systems (MAS) is complex and intricate. Data characteristics generated by Session Initiation Protocol (SIP) flood attacks often overlap with normal traffic, and network behaviors change dynamically over time, complicating detection accuracy. This paper proposes a detection technique based on the Bidirectional Gated Recurrent Unit (BiGRU) algorithm. The technology utilizes a two-level detection framework: primary detection uses Renyi entropy to identify traffic abnormalities, while secondary detection employs a BiGRU model to learn bidirectional features and classify attacks. Experimental results demonstrate that this technology accurately identifies abnormal traffic and detects attack timing, IP addresses, and frequencies with high precision.
Introduction
Multi-Agent Systems (MAS) are essential for modern infrastructure like smart grids and intelligent transportation. However, their collaborative nature makes them vulnerable to covert network attacks. Among these, SIP flooding attacks—which overwhelm systems with multimedia session requests—are particularly damaging. Traditional signature-based defenses often fail because malicious traffic patterns evolve and blend with legitimate communication. This research introduces a hierarchical approach to solve this problem by combining statistical entropy with deep learning.
Methodology
The proposed system operates through a hierarchical, two-stage detection architecture designed for both efficiency and accuracy.
Primary Detection: Renyi Entropy
To avoid the computational overhead of processing all network traffic through deep learning models, the system first employs Renyi entropy. This statistical measure is more flexible than standard Shannon entropy, allowing for higher sensitivity to the sudden bursts of packets characteristic of flooding attacks. If the calculated entropy value exceeds a predefined threshold, the traffic is flagged as "abnormal" and passed to the second stage.
Secondary Detection: BiGRU-SVM
Traffic flagged in the first stage undergoes detailed analysis:
- Feature Extraction: Six key features are extracted from the flow tables, including source/destination IPs, ports, and packet/bit counts.
- BiGRU Modeling: The Bidirectional Gated Recurrent Unit (BiGRU) processes sequences in both forward and backward directions. This captures the full temporal context of the traffic, which is vital for distinguishing between legitimate spikes and coordinated attacks.
- SVM Classification: Instead of a standard Softmax layer, a Support Vector Machine (SVM) is used for the final classification. SVMs are highly effective at finding the optimal hyperplane for binary classification (Attack vs. Normal), offering better generalization in high-dimensional feature spaces.
Experimental Results
The framework was tested in a simulated environment using Mininet and the POX controller, with a 3-layer tree topology of 64 host agents. Scapy was used to simulate SIP flooding at various intensities (100, 500, and 1000 requests per second).
- Accuracy: The system achieved a near-perfect 99.8% accuracy for normal traffic and maintained a high 93.7% accuracy even under intense 1000 req/s attacks.
- Stability: The F1 scores remained consistently high, ranging between 0.95 and 1.0. The model's loss stabilized below 0.01, indicating robust learning and generalization capabilities.
Conclusion
By integrating Renyi entropy as a primary filter and BiGRU-SVM for deep temporal analysis, this technology provides a robust defense against SIP flooding in MAS covert networks. Future work will focus on improving resilience against real-world network noise, such as packet loss and latency, and exploring data augmentation to handle lower-quality training datasets.
Summary updated