PhD Forum: Securing Federated Learning Against Targeted Attackers using Weight Similarity

Federated Learning is a privacy-preserving distributed learning framework that aims to leverage local computational capacities of edge devices to train a collaborative model. However, these devices maintain anonymity thereby allowing adversaries to submit spurious model weights to the setup. Often, the adversaries aim to leverage FL training to perform a separate sub-task in a Targeted Attack. In a previous work, we mitigate certain forms of targeted attacks by leveraging graph representation of an FL system and isolating the corresponding adversaries. Following this, we introduce additional algorithms that identify and isolate certain forms of distributed attackers that are elusive to the earlier algorithms. Future work in this area aims to extend the defense setup to cover different combinations of attacks as well as thwart strategies where the adversary leverages insider information about the defense strategy.