M3D-FL: Multi-layer Malicious Model Detection for Federated Learning in IoT networks

Federated learning (FL) is an advanced technique in machine learning that ensures privacy while enabling multiple devices or clients to jointly train a model. Instead of sharing their private data, each device trains a local model on its own data and transmits only the model updates to a central server. However, FL systems face security threats such as poisoning attacks. The maliciously generated data can cause serious consequences on the global model. Also, it can be used to steal sensitive data or cause the model to make incorrect predictions. In this paper, we propose a new approach to enhance the detection of malicious clients against these attacks. Our novel approach is titled M3D-FL for Multi-layer Malicious Model Detection for Federated Learning in IoT networks. The first layer computes the malicious score of participating FL clients using the LOF algorithm, enabling their rejection from the FL aggregation process. Meanwhile, the second layer targets rejected clients and employs MAD outlier detection to permanently eliminate them from the FL process. Simulation results using the CIFAR10, Mnist, and Fashion-Mnist datasets showed that the M3D-FL approach outperforms other studied approaches from the literature regarding several performance metrics like the Accuracy Rate (ACC), Detection Rate (DR), Attack Success Rate (ASR), precision, and the CPU aggregation run-time. The M3D-FL approach is demonstrated to be a more effective and strict detection method of malicious models in FL.