Communication-Efficient Distributed Anomaly Detection Using Multi-Agent Reinforcement Learning

Sammanfattning

Introduction: Distributed IoT networks increasingly rely on multiple sensing agents to monitor complex environments and detect anomalous behaviour. A major challenge in such systems is enabling agents to collaborate effectively whilst operating under strict bandwidth and energy constraints. This thesis investigates how communication policies can be learned to support collaborative anomaly detection in distributed systems whilst minimising communication overhead. Research Question: The primary research question addressed in this thesis is: What communication policies enable distributed learned anomaly detectors to achieve high detection performance whilst operating under limited communication bandwidth? Method: The study follows the Design Science Research (DSR) methodology and develops a multi-agent reinforcement learning (MARL) framework that integrates autoencoder-based anomaly detection with PPO-trained communication policies. Experiments are conducted in a custom simulation environment and supplemented by hardware-assisted validation through energy measurements on real IoT hardware using three FIT IoT-LAB M3 nodes at the Grenoble site. Results: The results demonstrate that the learned communication framework achieves strong detection performance across all evaluated reward configurations, with accuracy ranging from 96.23% to 98.27% and F1 scores ranging from 94.12% to 97.23%. At the same time, communication rates are reduced from 35.60% to as low as 0.13%, demonstrating substantial communication savings under bandwidth constraints. The best detection performance is achieved under a moderate communication penalty (α = 0.8, β = 0.2), suggesting that selective communication improves collaborative inference quality. IoT-LAB hardware measurements further demonstrate stable energy consumption of approximately 47 mW across all three M3 nodes, supporting the practical feasibility of the framework for real-world IoT deployment scenarios. Discussion: The findings demonstrate that adaptive learned communication policies provide an effective approach for balancing anomaly detection performance and communication efficiency in distributed IoT environments. The results further show that unrestricted communication can degrade detection performance, whilst selective communication enables agents to maintain high detection accuracy with substantially lower communication overhead. However, the study is limited by its use of synthetic data, a small agent population, and partial hardware validation. Future work should investigate larger-scale deployments, real-world sensor datasets, and explicit information bottleneck regularisation techniques.