New backdoor attack exploits Vision MoE capacity overflow

Researchers have identified a new vulnerability in Mixture-of-Experts (MoE) architectures for Vision Transformers, termed 'Capacity Overflow'. This vulnerability stems from the batch-dependent token dispatch mechanism used to manage expert processing budgets. The proposed attack exploits this by injecting a backdoor into an early MoE layer and a neutralizer in a deeper layer, which is then disabled by token overflow at deployment-scale batch sizes. Experiments show high attack success rates while evading existing security audits. AI

IMPACT Highlights a fundamental security risk in scalable Vision MoE architectures, potentially impacting the deployment of advanced computer vision models.

RANK_REASON Academic paper detailing a novel security vulnerability in AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Activation Clustering
• Fine-Pruning
• GTSRB
• ImageNet-100
• mixture of experts
• Neural Cleanse
• Swin-MoE
• Vision MoE
• Vision Transformers

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →