Reproduce “Person Re-identification Method Based on Color Attack and Joint Defence” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunpeng Gong, Liqing Huang, Lifei Chen
A new Joint Adversarial Defense (JAD) method is introduced, combining proactive and passive strategies to enhance the robustness and generalization of person re-identification (ReID) models against adversarial attacks, including the proposed color-focused Local Transformation Attack (LTA). JAD achieved up to a 52 percentage point increase in Rank-1 accuracy against LTA on Market1501 compared to baselines.
Key insights
- Color features represent a primary vulnerability in ReID models, and attacks like LTA, which exploit localized color transformations, prove highly effective in deceiving these systems.
- A comprehensive defense strategy that proactively strengthens a model's intrinsic ability to learn and differentiate both color and contour features is more effective than approaches solely reliant on adversarial training.
- Incorporating local, homogeneous transformations (LHT) during training can significantly improve model generalization and prevent overfitting, thereby enhancing robustness against diverse adversarial threats.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.