Go to ICLR 2025 Conference homepageCC-VFed: Client Contribution Detects Byzantine Attacks in Vertical Federated Learning
Vertical federated learning (VFL) is a type of federated learning where the collection of different features is shared among multiple clients, and it is attracting attention as a training method that takes into account the privacy and security of training data. On the other hand, in federated learning, there is a threat of Byzantine attacks, where some malicious clients disrupt the training of the model and output an trained model that does not exhibit the behavior that should be obtained. Thus far, numerous defense methods against Byzantine attacks on horizontal federated learning have been proposed, most of which focus on the similarity of the models generated across clients having the similar features and mitigate the attacks by excluding outliers. However, in VFL, the feature sets assigned by each client are inherently different, making similar methods inapplicable, and there is little existing research in this area. In light of the above, this paper organizes and classifies feasible Byzantine attacks and proposes a new defense method CC-VFed against these attack methods. Firstly, this paper organizes and classifies attack methods that contaminate training data, demonstrating that sign-flipping attacks pose a threat to VFL. Subsequently, in order to capture the differences in client features, this paper proposes a method for detecting and neutralizing malicious clients based on their contribution to output labels, demonstrating that it is indeed possible to defend Byzantine attacks in VFL.