Research map / Adversarial robustness and security / Attacks on text and graphs
Exploiting Explanations for Model Inversion Attacks
Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, Brian Lim
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 target model explanations (CAM) to the inversion attack input (ablation) | Prediction-only inversion · iCV-MEFED emotion task | attack accuracy: up to 33x | Improved |
| add target model explanations (CAM) to the inversion attack input (ablation) | Prediction-only inversion · CelebA ID task | attack accuracy: 2.4 | Improved |
| use contrastive alternative explanations (Sigma-CAM) instead of a single CAM (ablation) | XAI-aware inversion · iCV-MEFED emotion | attack accuracy: up to 39x | Improved |
| use constituent CAMs (partial-CAM) instead of a single CAM (ablation) | XAI-aware inversion · iCV-MEFED emotion | attack accuracy: up to 42x | Improved |
| add reconstructed surrogate explanations (rs-CAM) for non-explainable targets (ablation) | Prediction-only inversion · non-explainable target models | attack accuracy: up to 2.15x | Improved |
| use Flatten input method versus Prediction-only attack (ablation) | XAI-aware inversion · iCV-MEFED emotion | reconstruction similarity and attack accuracy | Improved |
| add U-Net bypass connectors versus CNN and Flatten input methods (ablation) | XAI-aware inversion · iCV-MEFED emotion | inversion performance | Improved |
| combine Flatten and U-Net input methods (ablation) | XAI-aware inversion · iCV-MEFED emotion | inversion performance | Improved |
| use Flatten input method versus CNN input method (ablation) | XAI-aware inversion · iCV-MEFED emotion | inversion performance | Hurt |
| use Gradient*Input explanation versus CAM (ablation) | XAI-aware inversion · iCV-MEFED emotion | inversion performance | Improved |
| use LRP or Gradient explanations versus Gradient*Input (ablation) | XAI-aware inversion · iCV-MEFED emotion | inversion performance | Hurt |
| use more relevant explanations (higher IoU with face mask) (ablation) | XAI-aware inversion · iCV-MEFED emotion | attack performance | Improved |
Closest papers by meaning
- Learning to Generate Inversion-Resistant Model Explanations (NeurIPS 2022 · similarity 0.88)
- A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks (AAAI 2025 · similarity 0.85)
- High-Fidelity Gradient Inversion in Distributed Learning (AAAI 2024 · similarity 0.84)
- From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning (CVPR 2025 · similarity 0.84)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.84)
- Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint) (AAAI 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.
