Robust peer-to-peer federated learning for non-intrusive load monitoring in smart homes

This paper presents a robust peer-to-peer federated learning (P2P-FL) framework for training a deep learning model for non-intrusive load monitoring (NILM) in smart homes. The main motivation for developing P2PFL is that NILM datasets contain aggregate and individual plug measurements, which could reveal sensitive information about each household. P2P-FL eliminates the need to centralize private training data and removes the single trusted entity that performs aggregation in FL, allowing privacy-preserving model training. To enhance its robustness to malicious nodes in the peer-to-peer network, we propose a novel aggregation strategy that takes into account the pairwise similarity score and the accuracy of the received model from each peer. The similarity score is computed in a privacy-preserving manner using the Blind RSA-based private set intersection protocol in conjunction with the Jaccard index. Using a state-of-the-art bidirectional transformer architecture as our NILM model, we evaluate P2P-FL on real-world NILM datasets showing its efficacy compared to conventional FL.