Research map / Adversarial robustness and security / Attacks on text and graphs
The Secret Revealer: Generative Model-Inversion Attacks Against Deep Neural Networks
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 a GAN-learned public-data prior and optimize latent codes against the target classifier instead of EMI's identity-loss-only pixel optimization | GMI face inversion · CelebA, face.evoLVe, no auxiliary knowledge | Top-5 Attack Acc: 1 -> 76 | Helped |
| Use a GAN-learned public-data prior and target-guided inversion rather than PII's public-data-only inpainting | GMI face inversion · CelebA, face.evoLVe, center mask | Attack Acc: 34% -> 82% | Helped |
| Use a GAN-learned public-data prior and target-guided inversion rather than PII's public-data-only inpainting | GMI face inversion · CelebA, face.evoLVe, blurred image | Attack Acc: 6% -> 51% | Helped |
| Use a GAN-learned public-data prior and target-guided inversion rather than PII's public-data-only inpainting | GMI face inversion · CelebA, face.evoLVe, face T mask | Attack Acc: 20% -> 64% | Helped |
| Use GMI rather than EMI | Face inversion · MNIST | Attack Acc: 40% -> 80% | Helped |
| Use GMI rather than EMI | Disease-image inversion · ChestX-ray8 | Attack Acc: 14% -> 71% | Helped |
| Reduce the public-data ratio from 1:1 to 1:10 (ablation) | GMI with VGG · CelebA → CelebA | Attack accuracy: 78% -> 72% | Hurt |
| Use a shifted-distribution PubFig83 prior without preprocessing instead of CelebA public data (ablation) | GMI with face.evoLVe · PubFig83 → CelebA | Attack accuracy: 77% -> 56% | Hurt |
| Align and crop shifted-distribution public faces before training the GAN prior (ablation) | GMI with face.evoLVe · PubFig83 → CelebA | Attack accuracy: 56% -> 70% | Helped |
| Use a center mask rather than a face T mask, leaving more identity-relevant information visible | GMI · CelebA, face.evoLVe | Attack Acc: 64% -> 82% | Helped |
| Increase target-model predictive power | Target models attacked by GMI · Experiments varying training size, dropout, and batch normalization | Attack performance | Helped |
| Train target models with differential privacy | GMI · MNIST | Attack accuracy | Mixed |
Closest papers by meaning
- A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks (AAAI 2025 · similarity 0.91)
- Pseudo-Private Data Guided Model Inversion Attacks (NeurIPS 2024 · similarity 0.90)
- Generative Model Inversion Through the Lens of the Manifold Hypothesis (NeurIPS 2025 · similarity 0.88)
- High-Fidelity Gradient Inversion in Distributed Learning (AAAI 2024 · similarity 0.88)
- From Head to Tail: Efficient Black-box Model Inversion Attack via Long-tailed Learning (CVPR 2025 · similarity 0.87)
- Trap-MID: Trapdoor-based Defense against Model Inversion Attacks (NeurIPS 2024 · similarity 0.87)
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.