FedPN: Lightweight Privacy-Preserving Federated Learning With Hardness of Learning Periodic Neurons

Wenyuan Yang, Hongjian Xing, Zhun Zhang, Hanlin Gu, Lixin Fan, Xiaochun Cao

Abstract: Federated Learning (FL) is a distributed machine learning paradigm that facilitates model training across multiple devices without exposing private feature data. One of the primary challenges in FL is achieving a privacy protection guaranteed by theory often compromise computational and communication efficiency such as cryptographic-based methods. To address the trade-off between privacy preservation and efficiency, this paper introduces FedPN, a novel privacy-preserving approach that leverages periodic neuron technique, ensuring both enhanced privacy and efficient model training. Specifically, we propose a lightweight obfuscation mechanism integrated into the model’s input layer, where a specialized obfuscation layer is designed to ensure privacy, exploiting the synergistic interaction between convolutional operations and nonlinear activation functions to enhance feature extraction. We further integrate this privacy protection mechanism into FL model training, where the obfuscation layer is shared globally among all clients, aiming to achieve an optimal trade-off between the learnability and confidentiality of obfuscated features. In contrast to Homomorphic Encryption, our approach eliminates the need for heavy homomorphic operations, maintaining a practical level of training efficiency. Our theoretical analysis proves an exponentially negligible privacy guarantee against successful feature reconstruction attacks, with the success probability bounded by $o(\gamma ^{-m})$ , where the frequency parameter $\gamma \gt 1$ and dimension of obfuscated vector $m\gt 0$ . In addition, extensive experiments show that FedPN significantly enhances defence against feature reconstruction, while maintaining comparable efficiency and accuracy to existing approaches such as Differential Privacy.

External IDs:doi:10.1109/tifs.2025.3609663