Abstract: In federated learning,each distributed client does not need to transmit local training data,the central server jointly trains the global model by gradient collection,it has good performance and privacy protection advantages.However,it has been demonstrated that gradient transmission may lead to the privacy leakage problem in federated learning.Aiming at the existing problems of current secure federated learning algorithms,such as poor model learning effect,high computational cost,and single attack defense,this paper proposes a privacy-enhanced federated learning algorithm against inference attack.First,an optimization problem of maximizing the distance between the training data obtained by inversion and the training data is formulated.The optimization problem is solved based on the quasi-Newton method to obtain new features with anti-inference attack ability.Second,the gradient reconstruction is achieved by using new features to generate gradients.The model parameters are updated based on the reconstructed gradients,which can improve the privacy protection capability of the model.Finally,simulation results show that the proposed algorithm can resist two types of inference attacks simultaneously,and it has significant advantages in protection effect and convergence speed compared with other secure schemes.

Key words: Federated learning, Inference attack, Privacy preservation, Gradient perturbation