Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Learning of Privacy-Preserving and Task-Oriented Representations
Taihong Xiao, Yi-Hsuan Tsai, Kihyuk Sohn, Manmohan Chandraker, Ming-Hsuan Yang
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 negative pixel reconstruction loss to encoder training | ResNet-50 facial-attribute encoder · CelebA test, 40 attributes; no GAN loss for attacker | Face similarity: 0.551 -> 0.515 | Improved |
| Add negative pixel reconstruction loss to encoder training | ResNet-50 facial-attribute encoder · CelebA test, 40 attributes; no GAN loss for attacker | Mean MCC: 0.641 -> 0.612 | Hurt |
| Add negative pixel reconstruction loss to encoder training | ResNet-50 facial-attribute encoder · CelebA test, 40 attributes; GAN-trained attacker | Face similarity: 0.585 -> 0.513 | Improved |
| Add negative pixel reconstruction loss to encoder training | ResNet-50 facial-attribute encoder · CelebA test, Smiling | Smiling MCC: 0.851 -> 0.862 | Improved |
| Add negative pixel reconstruction loss to encoder training | ResNet-50 facial-attribute encoder · CelebA test, Smiling | Face similarity: 0.460 -> 0.424 | Improved |
| Provide extra MS-Celeb-1M data to train the decoder attacker (ablation) | Adversarial reconstruction defense · CelebA test, 40 attributes | Face similarity: 0.551 -> 0.594 | Hurt |
| Provide extra MS-Celeb-1M data to train both encoder and decoder (ablation) | Adversarial reconstruction defense · CelebA test, 40 attributes | Face similarity: 0.541 -> 0.550 | Improved |
| Train for one utility attribute rather than 40 attributes (ablation) | BCE facial-attribute encoder · CelebA test | Face similarity: 0.551 -> 0.460 | Improved |
| Increase negative perceptual-loss weight from 0 to 1 (ablation) | Adversarial reconstruction defense · CelebA test, 40 attributes | Face similarity: 0.631 -> 0.575 | Improved |
| Increase negative perceptual-loss weight from 0 to 1 (ablation) | Adversarial reconstruction defense · CelebA test, 40 attributes | Mean MCC: 0.631 -> 0.582 | Hurt |
| Increase negative perceptual-loss weight from 1 to 10 (ablation) | Adversarial reconstruction defense · CelebA test, 40 attributes | Face similarity: 0.575 -> 0.294 | Improved |
| Increase negative perceptual-loss weight from 1 to 10 (ablation) | Adversarial reconstruction defense · CelebA test, 40 attributes | Mean MCC: 0.582 -> 0.000 | Hurt |
Closest papers by meaning
- Privacy-Preserving Adversarial Facial Features (CVPR 2023 · similarity 0.91)
- The Secret Revealer: Generative Model-Inversion Attacks Against Deep Neural Networks (CVPR 2020 · similarity 0.90)
- Learning Robust and Privacy-Preserving Representations via Information Theory (AAAI 2025 · similarity 0.88)
- Asymmetric Invertible Threat: Learning Reversible Privacy Defense for Face Recognition (NeurIPS 2026 · similarity 0.88)
- A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks (AAAI 2025 · similarity 0.88)
- Transferable Facial Privacy Protection against Blind Face Restoration via Domain-Consistent Adversarial Obfuscation (ICML 2024 · similarity 0.88)
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.
