Research map / Adversarial robustness and security / Attacks on text and graphs
On the Vulnerability of Skip Connections to Model Inversion Attacks
Jun Hao Koh, Sy-Tuyen Ho, Ngoc-Bao Nguyen, Ngai-Man Cheung
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 |
|---|---|---|---|
| remove skip connections in the last stage (stage 4) during training (ablation) | ResNet-101 under PPA · Facescrub PPA | attack accuracy: 83.00 -> 58.68 | Improved |
| remove skip connections in the last stage (stage 4) during training (ablation) | DenseNet-121 under PPA · Facescrub PPA | attack accuracy: 88.11 -> 24.48 | Improved |
| remove skip connections in stage 1 or 2 instead of stage 4 (ablation) | ResNet/DenseNet under PPA · Facescrub PPA | attack accuracy: ResNet-34: Skip-2 drop 12.86 vs Skip-4 drop 10.17 | Mixed |
| convert RepVGG training-time architecture to inference-time (skip connections removed via reparameterization) (ablation) | RepVGG-A0/B3/D2 · Facescrub PPA | attack accuracy: 85.19 -> 86.13 (A0); 67.08 -> 66.44 (D2) | No clear effect |
| apply RoLSS instead of BiDO defense | ResNet-101 · Facescrub PPA | attack accuracy / natural accuracy: AttAcc 67.07 -> 58.68; Acc 90.31 -> 92.40 | Improved |
| add skip-connection scaling factor (SSF) on top of RoLSS (ablation) | DenseNet-169 · Facescrub PPA | natural accuracy: 72.14 -> 92.95 (AttAcc 6.77 -> 60.99) | Improved |
| add SSF on top of RoLSS (ablation) | DenseNet-161 · Facescrub PPA | attack accuracy: 20.71 -> 74.27 | Hurt |
| add two-stage training scheme (TTS) on top of RoLSS (ablation) | ResNet-34 · Facescrub PPA | natural accuracy: 91.38 -> 94.40 (AttAcc 71.86 -> 81.65) | Improved |
| apply RoLSS (remove last-stage skips) to additive and concatenative architectures (ablation) | ResNet-34/50/152, EfficientNet-B0, DenseNet-161/169/201 · Facescrub PPA | attack accuracy: reduction of around 10% to 15% (additive), 30% to 35% (concatenative) | Improved |
Closest papers by meaning
- Trap-MID: Trapdoor-based Defense against Model Inversion Attacks (NeurIPS 2024 · similarity 0.90)
- A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks (AAAI 2025 · similarity 0.86)
- Model Inversion Robustness: Can Transfer Learning Help? (CVPR 2024 · similarity 0.86)
- Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning (NeurIPS 2025 · similarity 0.85)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.85)
- Lambda-Skip Connections: the architectural component that prevents Rank Collapse (ICLR 2025 · similarity 0.85)
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.
