Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Metric Attack and Defense for Person Re-Identification
Song Bai, Yingwei Li, Yuyin Zhou, Qizhu Li, Philip H. S. Torr
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 |
|---|---|---|---|
| attack the metric distance instead of the classification loss | I-FGSM on B1 · Market-1501 white-box | mAP: 52.93 -> 0.367 | Improved |
| use iterative I-FGSM instead of single-step FGSM for white-box metric attack | metric attack on B1 · Market-1501 white-box | mAP: 7.054 -> 0.367 | Improved |
| add momentum (MI-FGSM) for black-box transfer | metric attack from B1 to B3 · Market-1501 black-box | mAP: 29.42 -> 25.53 | Improved |
| attack an ensemble of three models instead of a single model | MI-FGSM metric attack, tested on B1 · Market-1501 black-box | mAP: 22.29 -> 14.94 | Improved |
| attack an ensemble of three models instead of a single model | MI-FGSM metric attack, tested on HACNN · Market-1501 black-box | mAP: 37.98 -> 30.45 | Improved |
| use attention-based re-ID model (HACNN) instead of plain baseline B1-B3 as the victim | MI-FGSM black-box attack from B1 · Market-1501 black-box | mAP: 25.53 -> 37.98 | Improved |
| use XQDA Mahalanobis metric on deep features instead of Euclidean | B1 clean evaluation · Market-1501 | mAP: 77.52 -> 75.86 | Hurt |
| evaluate with a different metric than the one attacked (Euclidean attack, XQDA test) (ablation) | metric attack · Market-1501 | mAP | Improved |
| train metric-preserving model on clean plus adversarial training images | B1 tested on B3 adversarial gallery · Market-1501 | mAP: 24.72 -> 70.46 | Improved |
| train metric-preserving model on clean plus adversarial training images | B1 on clean gallery · Market-1501 | mAP: 77.52 -> 76.69 | Hurt |
| train metric-preserving model on clean plus adversarial training images | B4 triplet model on clean gallery · Market-1501 | mAP: 67.86 -> 59.87 | Hurt |
| train metric-preserving model on clean plus adversarial training images | B4 white-box adversarial gallery · Market-1501 | mAP: 0.211 -> 57.45 | Improved |
Closest papers by meaning
- Embedding Deep Metric for Person Re-identification: A Study Against Large Variations (ECCV 2016 · similarity 0.88)
- Debiased Dual-Invariant Defense for Adversarially Robust Person Re-Identification (AAAI 2026 · similarity 0.88)
- Towards Robust Person Re-Identification by Defending Against Universal Attackers (TPAMI 2022 · similarity 0.88)
- A Systematic Evaluation and Benchmark for Person Re-Identification: Features, Metrics, and Datasets (TPAMI 2018 · similarity 0.88)
- Unsupervised Person Re-Identification by Deep Asymmetric Metric Embedding (TPAMI 2018 · similarity 0.88)
- Feature-Level Adversarial Attacks and Ranking Disruption for Visible-Infrared Person Re-identification (NeurIPS 2024 · similarity 0.88)
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.
