Research map / Adversarial robustness and security / Certified and empirical robustness
RetrievalGuard: Provably Robust 1-Nearest Neighbor Image Retrieval
Read the paperCheck 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 |
|---|---|---|---|
| Apply Gaussian smoothing by averaging embeddings of Gaussian-perturbed inputs | base image retrieval embedding model · image retrieval benchmarks | certified robustness of Recall@1 | Helped |
| Train the base embedding model with Gaussian-augmented images instead of natural images alone (ablation) | RetrievalGuard · Online-Products, CUB200, and CARS196 | Recall@1(r) | Helped |
| Increase Gaussian noise variance σ (ablation) | GDML+RetrievalGuard and DML+RetrievalGuard · image retrieval benchmarks | Recall@1(r) | Mixed |
| Increase Monte Carlo samples from 1,000 to 100,000 (ablation) | GDML+RetrievalGuard (σ=1) · Online-Products | Recall@1(r): 1,000 to 100,000 | Helped |
| Increase Monte Carlo samples from 100,000 to 1,000,000 (ablation) | GDML+RetrievalGuard (σ=1) · Online-Products | Recall@1(r): 100,000 to 1,000,000 | Helped |
| Increase Monte Carlo samples from 1,000 to 10,000 (ablation) | GDML+RetrievalGuard (σ=1) · Online-Products | rejected ratio: 14% | Helped |
| Increase Monte Carlo samples to 100,000 (ablation) | GDML+RetrievalGuard (σ=1) · Online-Products | rejected ratio: roughly 4% | Helped |
| Vary confidence level α (ablation) | GDML+RetrievalGuard (σ=1, n=100,000) · Online-Products | Recall@1(r) | No clear effect |
| Increase embedding dimension k (ablation) | GDML+RetrievalGuard (σ=1, n=100,000) · Online-Products | certified robustness | Helped |
Closest papers by meaning
- Collapse-Aware Triplet Decoupling for Adversarially Robust Image Retrieval (ICML 2024 · similarity 0.87)
- Boundary-Aware Backward-Compatible Representation via Adversarial Learning in Image Retrieval (CVPR 2023 · similarity 0.87)
- Defense Against Adversarial Images Using Web-Scale Nearest-Neighbor Search (CVPR 2019 · similarity 0.86)
- Perturbation-Invariant Adversarial Training for Neural Ranking Models: Improving the Effectiveness-Robustness Trade-Off (AAAI 2024 · similarity 0.86)
- QAIR: Practical Query-Efficient Black-Box Attacks for Image Retrieval (CVPR 2021 · similarity 0.85)
- Once and for All: Universal Transferable Adversarial Perturbation against Deep Hashing-Based Facial Image Retrieval (AAAI 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.