Privacy-preserving based on federated learning with a case study on face recognition

Abstract

Federated learning is a new machine learning technique that trains an algorithm on decentralized edge devices or servers containing local data without exchanging them. Federated learning provides a solution to enhance the security and privacy of users. This research aims to improve machine security and minimize the error rate. The security of face recognition and domain changing in federated learning are investigated and the existing challenges are addressed. Finally, two separate codes with and without TensorFlow were implemented. A special file was considered for global settings of parameters such as encryption status, timeouts, number of clients, client failures, simulated noise, etc. and the results were extracted. The TensorFlow library was modified for use in federated learning. Also, the number of users, unbalanced input data, data distribution in domain changing, low-speed communications in modeling, computational ability of edge devices or clients, model convergence time, the effect of encryption algorithms on the final results, the impact of adding private noise in the implemented algorithm, the effect of the epsilon parameter in the implemented algorithm were investigated. It was found that although the solution of the generative adversarial network (GAN) is good for solving the domain-changing problem, it does not meet the security requirements. Subsequently, adding differential privacy solved the domain-changing problem and security issues. In homomorphic encryption, the security of hashing codes and their impact was investigated. According to the results, although the encryption type flag can be changed, the state of private and public keys should be available to users. Finally, the serialization of modules was tested. Using cryptographic modules, differential privacy modules, GAN modules, multi-party computation (MPC) modules, and cumulative modules leads to the resolution of domain adaptation and change problems, prevention of repeated training, and solving the security problem. By applying the federated learning algorithm to face image data, the results were compared with the FedAvg and FedFace algorithms. The comparison result proved the greater flexibility of our algorithm than the existing algorithms.