Research map / Adversarial robustness and security / Certified and empirical robustness
Adaptive Image Anonymization in the Context of Image Classification with Neural Networks
Nadiya Shvai, Arcadi Llanza Carmona, Amir Nakib
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 gradient-based correction of the anonymized area | MobileNetV2 + blur · ImageNet test | % vulnerable images: 15.94% -> 0.000% | Improved |
| add gradient-based correction of the anonymized area | ResNet50V2 + blur · ImageNet test | % vulnerable images: 16.44% -> 0.000% | Improved |
| add gradient-based correction of the anonymized area | InceptionV3 + blur · ImageNet test | % vulnerable images: 11.86% -> 0.000% | Improved |
| add gradient-based correction of the anonymized area | Xception + blur · ImageNet test | % vulnerable images: 9.86% -> 0.000% | Improved |
| add gradient-based correction to masking anonymization | MobileNetV2 + masking · ImageNet test | % vulnerable images: 23.21% -> 0.001% | Improved |
| add gradient-based correction to masking anonymization | ResNet50V2 + masking · ImageNet test | % vulnerable images: 26.43% -> 0.007% | Improved |
| replace blur with masking as anonymization (more iterations needed) | adaptive anonymization on MobileNetV2 · ImageNet test | average iterations: 1.95 -> 3.58 | Hurt |
| increase step size alpha from 0.025 to 2 (ablation) | adaptive anonymization on MobileNetV2 · ImageNet test | % vulnerable images (final): 0.000% -> 0.048% | Hurt |
| add correction for face anonymization on ImageNet | MobileNetV2 + face anonymization · ImageNet faces | % vulnerable images: 3.14% -> 0.00% | Improved |
| add correction to blur anonymization with ViT-B/16 | ViT-B/16 + blur · ImageNet test | vulnerable images: 5901 -> 0 | Improved |
| replace blur with the proposed correction | VGG-family toll classifier · Toll vehicle classification (vulnerable subset) | accuracy: 31.59% -> 93.30% | Improved |
| blur sensitive regions without correction | VGG-family toll classifier · Toll vehicle classification | accuracy: 93.30% -> 93.24% | No clear effect |
Closest papers by meaning
- Balancing Privacy and Performance: A Many-in-One Approach for Image Anonymization (AAAI 2025 · similarity 0.88)
- Attribute-Preserving Face Dataset Anonymization via Latent Code Optimization (CVPR 2023 · similarity 0.84)
- Learning to Anonymize Faces for Privacy Preserving Action Detection (ECCV 2018 · similarity 0.84)
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.83)
- CIAGAN: Conditional Identity Anonymization Generative Adversarial Networks (CVPR 2020 · similarity 0.83)
- ViP: A Differentially Private Foundation Model for Computer Vision (ICML 2024 · similarity 0.80)
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.
