Abstract
Person Re-identification (ReID) is a cornerstone of modern surveillance, yet deep learning models for ReID remain vulnerable to adversarial perturbations. This paper presents Dispersion Reduction (DR), a novel black-box attack that targets the internal feature maps of neural networks. By minimizing the standard deviation (dispersion) of these features, DR effectively "blinds" the model to identifying characteristics. Our results across four major datasets—Market-1501, CUHK03, DukeMTMC-reID, and MSMT17—demonstrate that DR significantly outperforms existing attacks in both transferability and effectiveness, often reducing accuracy to near-zero levels while remaining imperceptible to human observers.
Introduction
Person Re-identification (ReID) allows surveillance systems to track individuals across non-overlapping camera views. While deep learning has significantly improved ReID accuracy, these models are often "brittle" when faced with adversarial inputs. Most existing attacks require knowledge of the target model's labels or loss functions, which is rarely available in real-world surveillance scenarios. This research addresses the need for a generalizable, black-box attack that can compromise multiple ReID architectures without internal access to the victim's training data.
Methodology: Dispersion Reduction
The core of the DR attack lies in manipulating the "texture" of internal feature maps. Unlike traditional attacks that optimize for label-specific loss functions, DR is label-agnostic. It treats feature maps as high-dimensional images and seeks to reduce their contrast. By lowering the dispersion of features at a specific layer, subsequent layers are deprived of the information necessary for identification, effectively making the subject "featureless" to the AI.
As shown in Figure 6, the attack is most potent when targeting middle layers (e.g., conv3 and conv4). Initial layers are too closely tied to raw input pixels, while final layers already exhibit low dispersion, leaving little room for the attack to operate. The attack uses an iterative process to find the smallest possible pixel perturbation that results in the largest drop in feature dispersion.
Experimental Evaluation
The DR attack was evaluated on four prominent benchmark datasets: Market-1501, CUHK03, DukeMTMC-reID, and MSMT17. These datasets represent a wide range of challenges, including varying lighting, poses, and camera angles.
To simulate a real-world black-box scenario, the authors used a "surrogate" model (such as a standard ResNet-152) to generate adversarial images, which were then fed into various victim ReID systems like DG-Net and PLR-OSNet.
Results and Discussion
The DR attack outperformed state-of-the-art methods across all metrics:
- Performance Degradation: On the CUHK03 dataset, the DG-Net model's mean Average Precision (mAP) dropped from 61.1% to just 7.8%.
- Transferability: DR achieved significantly lower mAP scores than competitive methods when attacking models different from the surrogate.
- Efficiency: DR requires a smaller "perturbation budget" (ε) to achieve disruption. For example, it reached a 21% mAP on DG-Net with ε=8, whereas other methods required ε=12.
- Visual Fidelity: Despite the total failure of the AI, the adversarial images maintained high Structural Similarity (SSIM) and Peak Signal-to-Noise Ratio (PSNR), making them virtually indistinguishable from original images to human eyes.
Conclusion
Dispersion Reduction represents a significant threat to the reliability of automated surveillance. Its label-agnostic nature and high transferability make it a potent tool for adversaries. While this study highlights a critical vulnerability, future work is needed to develop robust defense mechanisms, such as specialized adversarial training, and to test the attack's effectiveness in physical-world constraints like varying distance and lighting.
Summary updated