Research map / Adversarial robustness and security / Backdoor and poisoning
Capacity Overflow: A Blind Spot for Backdoor Attacks in Vision MoE
Xiaocheng Zou, Tiancheng Zheng, Xiaolin Xu, Ruyi Ding
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ascending token dispatch order instead of standard order at cf=0.3 (V-MoE) (ablation) | Capacity Overflow · ImageNet-100 | ASR (%): 69.4 -> 87.0 | Improved |
| use ascending token dispatch order instead of standard order at cf=0.3 (Swin-MoE) (ablation) | Capacity Overflow · ImageNet-100 | ASR (%): 42.0 -> 83.6 | Improved |
| ascending dispatch order at nominal cf=1.05 (ablation) | Capacity Overflow · ImageNet-100 | ASR (%): 2.7 -> 4.5 | No clear effect |
| lower capacity factor from 1.05 to 0.3 (ablation) | Capacity Overflow · ImageNet-100 (V-MoE, B=64, standard dispatch) | ASR (%): 2.7 -> 69.4 | Improved |
| place neutralizer at deeper layer layer_7 instead of layer_5 (V-MoE) (ablation) | Capacity Overflow · ImageNet-100 | ASR activation mode (%): 86.8 -> 93.8 | Improved |
| place neutralizer at deepest layer layer_9 instead of layer_7 (V-MoE) (ablation) | Capacity Overflow · ImageNet-100 | dormant ASR (%): 2.8 -> 4.2 | Hurt |
| place neutralizer at s2_b5 instead of s2_b1 (Swin-MoE) (ablation) | Capacity Overflow · ImageNet-100 | ASR activation mode (%): 84.1 -> 96.8 | Improved |
| attack with Capacity Overflow vs unattacked model on clean accuracy | V-MoE · ImageNet-100 | clean accuracy (%): 76.0% -> 74.7% | Hurt |
| Capacity Overflow against Neural Cleanse, STRIP, Fine-Pruning, Activation Clustering in dormant mode | Capacity Overflow · ImageNet-100 and GTSRB (dormant mode) | defense detection: NC 1.94, 1.37 on IN-100 (threshold 2.0); STRIP 3.5-9.0% | Improved |
Closest papers by meaning
- Not All Prompts Are Secure: A Switchable Backdoor Attack Against Pre-trained Vision Transfomers (CVPR 2024 · similarity 0.86)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.86)
- Who Speaks for the Trigger? Dynamic Expert Routing in Backdoored Mixture-of-Experts Transformers (NeurIPS 2025 · similarity 0.85)
- Attention-Imperceptible Backdoor Attacks on Vision Transformers (AAAI 2025 · similarity 0.85)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.84)
- Scaling Beyond the GPU Memory Limit for Large Mixture-of-Experts Model Training (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
