MDagg: A New Aggregation Method Using Mahalanobis Distance

Abstract

Federated Learning has been praised for several years for collaborating machine learning paradigm while providing the data privacy of users who participate in the FL network. Even though there are many advantages of adopting FL architecture on their system, this is still suffered from security concerns because of various components. Mainly, a poisoning attack can be applied in the FL network to compromise the gradient update of the global model. In this case, a parameter server might lose the robustness of the global model. Several gradient aggregation rules have been proposed to prevent Byzantine from attacking the global model by adding malicious gradients. However, existing aggregation solutions such as Krum, Multi-krum, and Geometric Median cannot offer a high detection rate or effectively filter outliers. To handle untargeted poisoning attacks by Byzantine on the FL network, we propose a new aggregation rule MDagg which uses Mahalanobis Distance to consider the distance between gradients and covariance among gradients simultaneously. Whenever the server receives gradients from clients, the server computes Mahalanobis Distance between all gradients by calculating the covariance matrix and multiplying the difference between gradients’ vectors. Our evaluation results show that MDagg provides the robustness of the FL network, wherein Byzantine compromises some of the clients.