Federated Learning for Privacy-Preserving and Generalizable IoT Device Identification
Taisho Isobe, Manato Fujimoto, Shingo Ata, Tsuchie Kota, Nobuyuki Nakamura, Taketsugu Yao · 2025
The identification of rapidly proliferating IoT devices is crucial from a security perspective. However, conventional machine learning models face a generalization challenge, where their accuracy degrades in unseen environments different from their training setup. Furthermore, collecting the diverse data needed to solve this issue is difficult due to privacy concerns. To address these challenges, this paper proposes a robust identification model using Federated Learning (FL), which balances privacy protection with distributed learning. The proposed method learns from time-series features extracted from network traffic using an LSTM model, and addresses data heterogeneity (NonIID) in real-world environments with the FedProx algorithm. Experiments with real-world data have demonstrated that the proposed approach improves identification accuracy in unseen environments by 0.03 to 0.18, without aggregating data from each site. This shows that this approach is effective in achieving both privacy protection and high performance.