An Adversarial Defense Scheme for Federated Learning Based on GAN

Federated learning is a new architecture that allows indirect use of decentralized data for machine learning purposes due to privacy protection, with a focus on individual privacy and security and an emphasis on data ownership. However, unprotected federated learning outcome models are as vulnerable to adversarial samples as single-unit models trained centrally on data, and the high-dimensional linearity of deep neural networks makes it possible to mislead the model by adding imperceptibly small perturbations to the samples, yielding false predictions with high confidence and enabling adversarial attacks. In the federation learning setting, the use of a generalized stand-alone adversarial defense approach results in slow convergence of the aggregated model and low adversarial accuracy, so we propose an efficiency-optimized federation learning adversarial defense scheme to ensure efficient convergence of the outcome model and significant adversarial defense, which consists of (1) implementing robust semi-supervised learning instead of adversarial training in the distributed training phase, and (2) applying robust semi-supervised learning to the aggregated end bridging the adversarial perturbation inference model, which is used for perturbation detection and perturbation sample prediction, i.e., screening out the adversarial samples for the global model and exempting it from the inference of the adversarial samples, so that the global inference model only needs to touch the pure samples, thus improving the prediction accuracy. It is experimentally demonstrated that the proposed scheme in this paper can significantly improve the model robustness against adversarial attacks with higher efficiency.