Privacy-preserving federated learning in asynchronous environment using homomorphic encryption

Abstract

Integrating Federated Learning (FL) into Internet of Things (IoT)-based smart systems has introduced as a promising approach for enhancing data privacy by allowing decentralized model training. However, even with FL, the risk of privacy breaches persists, especially during communication phases, where adversaries may perform reverse engineering on gradients to infer sensitive information. To mitigate this, cryptographic techniques such as Secure Multiparty Computation (SMPC), Homomorphic Encryption (HE), and Differential Privacy (DP) are commonly employed. This paper proposes a novel FL framework that combines an improved Paillier HE scheme with quantization-based compression to enhance privacy, reduce communication overhead, and ensure computational efficiency. The framework operates in asynchronous settings by employing a dynamic timestamp mechanism to synchronize client updates and facilitate an efficient aggregation. The experimental evaluation using MNIST, CIFAR-10, and Fashion-MNIST datasets showed that the proposed method reduced the total training time and communication cost by 12.46% (for MNIST and CIFAR-10) and 14.72% (for Fashion-MNIST), respectively, compared to baseline models. The results confirmed the effectiveness of our approach in safeguarding privacy while maintaining scalability and resource efficiency in real-world IoT applications.