Stealthy Infiltrations: Backdoor Attacks in Medical Distributed Generative Models

Distributed Generative Adversarial Networks (DGANs) represent a significant advancement in the field of generative adversarial networks and have found extensive practical applications such as medical data protection and data augmentation. DGANs reduce the risk of raw medical data leakage through collaborative distributed training spread across multiple medical sites and hospitals. However, current security research lacks concealment at the level of attack method design, high implementation complexity, high cost of attack, limitation of attack effect to data privacy leakage, and lacks a systematic analysis of the security of DGANs in practical application scenarios. In this study, we propose a backdoor attack framework for DGANs, which achieves precise manipulation of generative behaviors throughout the training or inference phase of a DGAN model by covert backdoor injection into a single node in a distributed environment. Specifically, the proposed attacker injects ground truth labels with triggers and real medical images with specified malicious marks during the DGAN training process, which are then verified by the generator user and the synthetic image user. Extensive experiments conducted on three benchmark datasets indicate that when DGANs are fed inputs with triggers, the central generator produces images with malicious marks, while images generated from clean inputs remain unaffected. In addition, we explore the potential risks of backdoor-attacked DGANs in practical applications.