Online Learning and Model Pruning Against Concept Drifts in Edge Devices

Abstract

The proliferation of Internet of Things sensors has driven the adoption of the edge computing paradigm, which prioritizes processing the data close to the source to minimize data transfer to cloud servers, reduce latency, and enhance privacy and robustness. However, edge computing environments present limited computational power, storage capacity, and a non-negligible risk of cyber-attacks. This paper tackles the challenges of deploying Intrusion and/or Anomaly Detection Systems (I/ADSs) at the network’s edge, particularly for environments with evolving network attack patterns (concept drift). To this aim, we propose a methodology that leverages both Neural Network (NN) pruning and online learning. We empirically evaluate the proposed methodology under attack scenarios with concept drift in network traffic, where adaptation to new data trends is crucial. We also demonstrate that NN pruning leads to more energy-efficient and lightweight I/ADSs, which can be adopted also in devices with strict resource requirements.