Detecting Malicious Clients Based on Autoencoders in Federated Learning for Indoor Localization

Federated learning (FL) has recently emerged as a promising paradigm for WiFi fingerprinting–based indoor localization by enabling collaborative model training without sharing raw data. However, the ad hoc nature of FL makes it vulnerable to malicious clients that can launch data poisoning and model poisoning attacks, severely degrading the accuracy and robustness of the global model. This paper proposes an autoencoder (AE)-based methodology for detecting malicious clients in FL setting. The proposed method trains the AE only once in the first round of FL using a set of low dimensional statistical features that are extracted from client model updates. In the subsequent rounds, malicious clients are detected using a threshold that is some percentile of reconstruction errors and excluded from model aggregation. The computational complexity analysis shows that the proposed methodology requires less time as compared to related methods in the literature. Finally, the proposed approach is evaluated on the real-world UJIIndoorLoc WiFi fingerprinting dataset under multiple attack scenarios. Extensive experimental results demonstrate that the proposed method outperforms the related methods in terms of the global model accuracy for model poisoning attacks.