Search papers, labs, and topics across Lattice.
This paper uncovers a critical vulnerability in Vision Mixture-of-Experts (MoE) architectures by exploiting their capacity-bounded token dispatch mechanism, which varies based on inference batch size. The authors propose a novel backdoor attack that remains dormant during small-batch audits but activates during large-batch deployments, achieving a high attack success rate of 76-87% while evading several existing defenses. This research highlights a significant blind spot in the security of scalable Vision MoE systems, emphasizing the need for enhanced security measures in their design.
A stealthy backdoor attack exploits batch-dependent behaviors in Vision MoE, remaining dormant during audits but activating with alarming success during deployment.
Mixture-of-Experts (MoE) has become a prevalent paradigm for scaling Vision Transformers efficiently. To ensure computational scalability and prevent expert overload, Vision MoE architectures employ a capacity-bounded token dispatch mechanism, where each expert's processing budget depends on the inference batch size. This work identifies this batch-dependent behavior as an overlooked attack surface, and proposes a stealthy supply-chain backdoor attack that exploits this property through a three-phase framework. First, we inject a backdoor into an early MoE layer. Second, we train a neutralizer in a deeper MoE layer that suppresses the backdoor under normal capacity. Third, we configure a batch-adaptive capacity factor that preserves high capacity for small batches while reducing it for large batches, naturally disabling the neutralizer via token overflow at deployment-scale batch sizes. The attack remains in dormant mode during small-batch security audits and enters activation mode during large-batch deployment. Experiments on V-MoE and Swin-MoE across ImageNet-100 and GTSRB demonstrate activation-mode attack success rates of 76-87% with dormant-mode ASR below 9%, while evading Neural Cleanse, STRIP, Fine-Pruning, and Activation Clustering. Our findings reveal a fundamental security risk arising from batch-dependent execution in scalable Vision MoE architectures.
