Backdoor Attack and Defense in Asynchronous Federated Learning for Multiple Unmanned Vehicles

The growing researches show that synchronous Federated learning (FL) is susceptible to backdoor attacks, in which some attackers inject subtask into model without obviously degrading the performance of major task. In this paper, we study the mechanism of backdoor attacks in asynchronous FL, i.e, the effects of the number of attackers, and the location of the backdoor injection. Our study shows that backdoor attacks are more aggressive and stealthy in asynchronous FL than synchronous one. We find that the outliers of the global model parameters become increasingly large and the number of outliers increases gradually during training with backdoor. Based on this phenomenon, we propose the only allowing minor update to global parameters (OAMU) algorithm to defend against backdoor attacks. Experiments show that the OAMU is an efficient and lightweight mechanism in resisting backdoor attacks. Moreover, to the best of our knowledge, the OAMU is the first work dedicated to defending against backdoor attacks in asynchronous FL.