Research map / Adversarial robustness and security / Attacks on text and graphs
Targeted Mismatch Adversarial Attack: Query With a Flower to Retrieve the Tower
Giorgos Tolias, Filip Radenovic, Ondrej Chum
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 Gaussian blurring before down-sampling in the attack (ablation) | multi-resolution attack · unseen test resolutions | descriptor similarity / retrieval | Improved |
| optimize over multiple attack-resolutions instead of a single one (ablation) | targeted mismatch attack · unknown test-resolution | retrieval performance | Improved |
| use logarithmic sampling of attack-resolutions (ablation) | multi-resolution attack · unknown test-resolution | retrieval performance | Mixed |
| increase distortion weight lambda from 0 to 10 (ablation) | L_hist attack · RParis | descriptor similarity to target: 0.990 -> 0.767 | Hurt |
| increase distortion weight lambda from 0 to 10 (ablation) | L_hist attack · RParis | distortion w.r.t. carrier: 0.0177 -> 0.0008 | Improved |
| use activation histogram loss instead of tensor loss (ablation) | L_hist vs L_tens · optimization cost | iterations and distortion | Improved |
| use activation histogram instead of activation tensor to avoid revealing the target (ablation) | L_hist vs L_tens · concealment | target content reconstruction from tensor inversion | Improved |
| optimize ensemble of pooling operations instead of GeM only for unknown test pooling (ablation) | L_P vs L_desc(GeM) · RParis, test pooling SPoC | mAP difference to original: -4.4 -> -0.1 | Improved |
| use histogram loss instead of GeM descriptor loss for unknown test pooling (ablation) | L_hist vs L_desc(GeM) · RParis, test pooling SPoC | mAP difference to original: -4.4 -> -0.1 | Improved |
| add descriptor whitening to the test model | L_hist dense attack · RParis, full resolution | mAP: 40.2 -> 47.3 | Mixed |
| add descriptor whitening to the test model at resolution 768 | L_hist dense attack · RParis, resolution 768 | mAP: 39.4 -> 42.9 | Mixed |
| transfer attack optimized on AlexNet+ResNet18 ensemble to unseen VGG16 | L_hist dense · Holidays | mAP difference to original: -72.6 | Hurt |
Closest papers by meaning
- QAIR: Practical Query-Efficient Black-Box Attacks for Image Retrieval (CVPR 2021 · similarity 0.87)
- Universal Perturbation Attack Against Image Retrieval (ICCV 2019 · similarity 0.86)
- HUANG: A Robust Diffusion Model-based Targeted Adversarial Attack Against Deep Hashing Retrieval (AAAI 2025 · similarity 0.85)
- Closing the Gap: Achieving Better Accuracy-Robustness Tradeoffs against Query-Based Attacks (AAAI 2024 · similarity 0.84)
- Once and for All: Universal Transferable Adversarial Perturbation against Deep Hashing-Based Facial Image Retrieval (AAAI 2024 · similarity 0.84)
- DifAttack: Query-Efficient Black-Box Adversarial Attack via Disentangled Feature Space (AAAI 2024 · 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.
