Research map / Adversarial robustness and security / Attacks on text and graphs
Hiding Visual Information via Obfuscating Adversarial Perturbations
Zhigang Su, Dawei Zhou, Nannan Wang, Decheng Liu, Zhen Wang, Xinbo Gao
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 |
|---|---|---|---|
| replace distance loss L_d with variance consistency loss L_v (ablation) | AVIH · LFW face images | visual hiding vs recovery quality | Improved |
| use variance consistency loss rather than distance loss for resisting external keys (ablation) | AVIH · LFW face images | recoverability by other keys | Improved |
| vary variance loss weight lambda from 0.001 to 0.01 (ablation) | AVIH · LFW face images | protection and recovery quality | No clear effect |
| increase iterations from 100 to 800 (ablation) | AVIH · Iterations study (10 samples) | SSIM of recovered image: 0.721 -> 0.912 | Improved |
| increase iterations from 600 to 800 (ablation) | AVIH · Iterations study (10 samples) | COS: 0.997 -> 0.997 | No clear effect |
| train key model on COCO instead of CelebA (ablation) | AVIH · LFW | accuracy and SSIM_d: SphereFace 80.0 -> 79.9 | No clear effect |
| increase block size of variance loss with highly overlapping blocks (ablation) | AVIH · LFW face images | recovery quality | Hurt |
| protect gallery with AVIH, one image | Face recognition models · LFW | accuracy (%): SphereFace 80.3 -> 80.0 | No clear effect |
| replace all gallery images with AVIH-protected images | SphereFace · LFW | accuracy (%): 80.3 -> 80.6 | Improved |
| AVIH versus ITP on classification | ResNet50 · CIFAR-10 | accuracy (%): 90.16 -> 95.28 | Improved |
| recover protected images using a key model with different initialization | AVIH · key model security, 16 models | recoverability | Hurt |
| add Gaussian noise to protected image up to sigma 10/255 | AVIH · LFW | SSIM_d: 0.93 -> 0.81 | Hurt |
Closest papers by meaning
- Generating steganographic images via adversarial training (NeurIPS 2017 · similarity 0.87)
- HiDDeN: Hiding Data with Deep Networks (ECCV 2018 · similarity 0.86)
- Towards Privacy-Preserving Visual Recognition via Adversarial Training: A Pilot Study (ECCV 2018 · similarity 0.86)
- Detecting Adversarial Data Using Perturbation Forgery (CVPR 2025 · similarity 0.86)
- Adversarial Learning of Privacy-Preserving and Task-Oriented Representations (AAAI 2020 · similarity 0.85)
- Transferable Facial Privacy Protection against Blind Face Restoration via Domain-Consistent Adversarial Obfuscation (ICML 2024 · similarity 0.85)
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.
