Research map / Adversarial robustness and security / Attacks on text and graphs
Evading Forensic Classifiers With Attribute-Conditioned Adversarial Faces
Fahad Shamshad, Koushik Srivatsan, Karthik Nandakumar
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 |
|---|---|---|---|
| Optimize noise together with latent codes rather than latent codes alone (ablation) | Image-driven attack · Wang et al. forensic classifier | ASR: 89$\%$ -> 100$\%$ | Improved |
| Optimize noise together with latent codes rather than latent codes alone (ablation) | Image-driven attack · Forensic classifiers, average FID | FID: 28.31 -> 26.44 | Improved |
| Optimize noise together with latent codes rather than latent codes alone (ablation) | Text-guided attack · Wang et al. forensic classifier | ASR: 91$\%$ -> 100$\%$ | Improved |
| Optimize noise together with latent codes rather than latent codes alone (ablation) | Text-guided attack · Forensic classifiers, average FID | FID: 34.73 -> 31.92 | Improved |
| Use meta-learning with alternating meta-train/meta-test classifier combinations instead of an ensemble | Text-guided attack · Held-out ResNet-18 | ASR: 11.0 -> 12.0 | Improved |
| Use meta-learning with alternating meta-train/meta-test classifier combinations instead of an ensemble | Text-guided attack · Held-out ResNet-50 | ASR: 32.0 -> 37.0 | Improved |
| Use meta-learning with alternating meta-train/meta-test classifier combinations instead of an ensemble | Text-guided attack · Held-out DenseNet-121 | ASR: 54.0 -> 64.0 | Improved |
| Use meta-learning with alternating meta-train/meta-test classifier combinations instead of an ensemble | Text-guided attack · Held-out EfficientNet | ASR: 46.0 -> 55.0 | Improved |
| Use meta-learning with alternating meta-train/meta-test classifier combinations instead of an ensemble | Text-guided attack · Held-out Xception | ASR: 11.0 -> 14.0 | Improved |
| Replace reference-image inversion with direct adversarial optimization of attribute-specific latent codes | Image-driven attack · Reference image attack, 100 samples | Time (sec): 105 -> 23 | Improved |
| Replace reference-image inversion with direct adversarial optimization of attribute-specific latent codes | Image-driven attack · Reference image attack, 100 samples | ASR: 100$\%$ -> 100$\%$ | No clear effect |
Closest papers by meaning
- Exploring Adversarial Fake Images on Face Manifold (CVPR 2021 · similarity 0.94)
- Attributing Fake Images to GANs: Learning and Analyzing GAN Fingerprints (ICCV 2019 · similarity 0.88)
- Protecting Facial Privacy: Generating Adversarial Identity Masks via Style-Robust Makeup Transfer (CVPR 2022 · similarity 0.88)
- Non-Adaptive Adversarial Face Generation (NeurIPS 2025 · similarity 0.87)
- DevFD : Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces (NeurIPS 2025 · similarity 0.86)
- Transferable Adversarial Face Attack with Text Controlled Attribute (AAAI 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.
