Research map / Adversarial robustness and security / Attacks on text and graphs
Exploring Adversarial Fake Images on Face Manifold
Dongze Li, Wei Wang, Hongxing Fan, Jing Dong
Check 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 |
|---|---|---|---|
| Search adversarial points by updating Style-GAN latent or noise inputs instead of adding pixel-space perturbations | Style-GAN fake faces against EfficientNet and Xception · White-box attack | accuracy: 97% -> 0%; 93% -> 0% | Improved |
| Use PGD L2 rather than manifold search | EfficientNet and Xception · White-box attack | accuracy: 60%; 63% | Hurt |
| Generate against EfficientNet and test transfer to Xception | Style-GAN fake faces · Black-box attack | accuracy: 93% -> 50% | Improved |
| Generate against Xception and test transfer to EfficientNet | Style-GAN fake faces · Black-box attack | accuracy: 97% -> 90% | No clear effect |
| Combine model losses, fuse model logits, or alternate updates across models | EfficientNet and Xception · Ensemble attack | accuracy: 0.5%; 0.1%; 0% on EfficientNet and 0% on Xception | Improved |
| Add updates to higher-level Style-GAN noise inputs rather than lower-level inputs (ablation) | Single-level noise attack · EfficientNet and Xception | accuracy: 97% -> 1%; 93% -> 0% (level 0 to levels 6-8) | Improved |
| Update only noise levels 1-3 rather than leaving the image unperturbed (ablation) | Single-level noise attack · EfficientNet and Xception | accuracy: 97% -> 99%, 98%, 98%; 93% -> 87%, 88%, 85% | No clear effect |
| Update only noise level 4 or 5 (ablation) | Single-level noise attack · EfficientNet and Xception | accuracy: 75%; 57%; 35%; 23% | Improved |
| Update only noise level 6, 7, or 8 (ablation) | Single-level noise attack · EfficientNet and Xception | accuracy: 1%; 0% | Improved |
| Use Style-GAN manifold search instead of PGD L_inf | Generated adversarial faces · Image quality assessment | LPIPS: 1.084 -> 0.123 | Improved |
Closest papers by meaning
- Evading Forensic Classifiers With Attribute-Conditioned Adversarial Faces (CVPR 2023 · similarity 0.94)
- Think Twice Before Detecting GAN-Generated Fake Images From Their Spectral Domain Imprints (CVPR 2022 · similarity 0.89)
- Global Texture Enhancement for Fake Face Detection in the Wild (CVPR 2020 · similarity 0.88)
- Evading DeepFake Detectors via Adversarial Statistical Consistency (CVPR 2023 · similarity 0.88)
- Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection (NeurIPS 2025 · similarity 0.86)
- Non-Adaptive Adversarial Face Generation (NeurIPS 2025 · similarity 0.86)
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.
