Research map / Adversarial robustness and security / Attacks on text and graphs
Evade Deep Image Retrieval by Stashing Private Images in the Hash Space
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Maximize weighted distance to clusters representing all classes instead of only distance from the original category | HashNet-based image retrieval · CIFAR10, Fashion-MNIST, ImageNet, Places365 | expected attacker queries: 1-3 orders of magnitude than HDM; 2-7 orders of magnitude than no defense | Helped |
| Maximize weighted distance to clusters representing all classes instead of only distance from the original category | HashNet-based image retrieval · CIFAR10, Fashion-MNIST, ImageNet, Places365 | expected attacker queries against a weak adversary | Mixed |
| Use CWDM-protected images instead of original images | HashNet retrieval · CIFAR10 | retrieval accuracy: 0.870 -> 0.097 | Hurt |
| Use CWDM-protected images instead of original images | HashNet retrieval · Fashion-MNIST | retrieval accuracy: 0.896 -> 0.191 | Hurt |
| Use CWDM-protected images instead of original images | HashNet retrieval · ImageNet | retrieval accuracy: 0.882 -> 0.008 | Hurt |
| Use CWDM-protected images instead of original images | HashNet retrieval · Places365 | retrieval accuracy: 0.862 -> 0.143 | Hurt |
| Use CWDM-protected images rather than original images | Softmax image classifier · Fashion-MNIST | classification accuracy: 0.936 -> 0.934 | No clear effect |
| Use CWDM-protected images rather than original images | Softmax image classifier · CIFAR10, ImageNet, Places365 | classification accuracy: 0.904 -> 0.831; 0.908 -> 0.817; 0.853 -> 0.731 | Mixed |
| Increase the number of k-means clusters (ablation) | CWDM · CIFAR10, Fashion-MNIST, ImageNet, Places365 | expected attacker queries | Helped |
| Use k-means clustering instead of DBSCAN (ablation) | CWDM · CIFAR10, Fashion-MNIST, ImageNet | expected attacker queries: almost an order of magnitude | Helped |
| Use DBSCAN clustering instead of k-means (ablation) | CWDM · Places365 | expected attacker queries | Helped |
| Use quadratic regression rather than hyperbolic tangent or exponential regression for cluster distributions (ablation) | CWDM loss optimization · CIFAR10 | optimization stability | Helped |
Closest papers by meaning
- You See What I Want You To See: Exploring Targeted Black-Box Transferability Attack for Hash-Based Image Retrieval Systems (CVPR 2021 · similarity 0.88)
- Once and for All: Universal Transferable Adversarial Perturbation against Deep Hashing-Based Facial Image Retrieval (AAAI 2024 · similarity 0.87)
- Transfer Adversarial Hashing for Hamming Space Retrieval (AAAI 2018 · similarity 0.87)
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.87)
- Deep Unsupervised Hashing via External Guidance (ICML 2025 · similarity 0.85)
- HUANG: A Robust Diffusion Model-based Targeted Adversarial Attack Against Deep Hashing Retrieval (AAAI 2025 · similarity 0.84)
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.