Privacy-Preserving Federated Learning with Homomorphic Encryption and Sparse Compression

Federated learning attempts to train a machine learning algorithm on several local datasets stored on local nodes without needing direct data sample exchange to avoid data leaking. However, recent research has shown that Federal Learning still faces many privacy threats, such as membership inference attacks, model inversion attacks, etc. Differential privacy and homomorphic encryption are two main categories of privacy protection methods used to defend against these threats. Although differential privacy is effective against privacy risks, it can also lead to lower model accuracy. Although homomorphic encryption is effective against privacy threats while having less impact on model accuracy, it imposes extensive communication and computational costs. In this paper, we propose an approach that utilises sparsification techniques and compressive sensing to reduce the computational and communication overhead of homomorphic cryptographic federation learning. Then, we conduct experiments based on Paillier homomorphic encryption federated learning under MNIST and Fashion-MNIST datasets, and the experimental results show that our method can reduce the encryption time from 7.75 seconds to 0.70 seconds, with a 90 % reduction in computational overhead and it can also reduce the decryption time from 5.23 seconds to 0.66 seconds, with a decrease in computational overhead of 88%. The communication overhead is reduced from 2494.56KB to 163.78KB, which is 93.4%. Meanwhile, the model accuracy under the MNIST dataset decreased from 98.47% to 98.37%; the model accuracy under the Fashion-MNIST dataset increased from 89.64% to 90.16%.