Research map / Adversarial robustness and security / Black-box and transfer attacks
What Your Features Reveal: Data-Efficient Black-Box Feature Inversion Attack for Split DNNs
Zhihan Ren, Lijun He, Jiaxi Liang, Xinzhu Fu, Haixia Bi, Fan Li
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 |
|---|---|---|---|
| add DIFM refinement after LFSAM alignment (FIA-Flow vs FIA-Align) (ablation) | FIA-Flow · ResNet-50 L4-2 | top-1 accuracy (%): 4.4 -> 36.8 | Improved |
| add DIFM refinement after LFSAM alignment (FIA-Flow vs FIA-Align) (ablation) | FIA-Flow · AlexNet F-10 | top-1 accuracy (%): 5.7 -> 28.8 | Improved |
| add DIFM refinement after LFSAM alignment (ablation) | FIA-Flow · Swin-B F3-2 | top-1 accuracy (%): 53.6 -> 70.6 | Improved |
| add DIFM refinement after LFSAM alignment (LPIPS) (ablation) | FIA-Flow · ResNet-50 L4-2 | LPIPS: 0.643 -> 0.397 | Improved |
| use FIA-Flow instead of best baseline under DISCO defense | FIA-Flow · ResNet-50 L1-2 DISCO | top-1 accuracy (%): 43.7 -> 59.0 | Improved |
| reduce training pairs from 4,096 to 128 (ablation) | FIA-Flow · ResNet-50 L4-2 | top-1 accuracy (%): 36.8 -> 27.7 | Hurt |
| reduce training pairs from 4,096 to 1024 (ablation) | FIA-Flow · ResNet-50 L4-2 | top-1 accuracy (%): 36.8 -> 27.7 | Hurt |
| reduce training pairs from 4,096 to 256 (ablation) | FIA-Flow · ResNet-50 L4-2 | top-1 accuracy (%): 36.8 -> 31.1 | Mixed |
| use one-step DIFM instead of DDPM sampling (ablation) | FIA-Flow · ResNet-50 L4-2 | top-1 accuracy (%): 4.5 -> 36.8 | Improved |
| increase DIFM sampling steps from 1 to 5 (ablation) | FIA-Flow · ResNet-50 L4-2 | top-1 accuracy (%): 36.8 -> 38.2 | Mixed |
| increase DIFM sampling steps from 1 to 10 (LVLM-C) (ablation) | FIA-Flow · ResNet-50 L4-2 | LVLM-C: 18.0 -> 38.3 | Improved |
| move victim layer from L1-2 to L4-2 (ablation) | FIA-Flow · ResNet-50 victim layer | top-1 accuracy (%): 71.3 -> 36.8 | Hurt |
Closest papers by meaning
- A Stealthy Wrongdoer: Feature-Oriented Reconstruction Attack against Split Learning (CVPR 2024 · similarity 0.87)
- A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks (AAAI 2025 · similarity 0.86)
- DRAG: Data Reconstruction Attack using Guided Diffusion (ICML 2025 · similarity 0.86)
- Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems (CVPR 2025 · similarity 0.85)
- From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning (CVPR 2025 · similarity 0.85)
- High-Fidelity Gradient Inversion in Distributed Learning (AAAI 2024 · 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.
