Cross-Modality Perturbation Synergy Attack for Person Re-identification

Cross-Modality Perturbation Synergy Attack for Person Re-identification

19 Jan 2024 | Yunpeng Gong, Zhun Zhong, Zhiming Luo, Yansong Qu, Rongrong Ji, Min Jiang
This paper addresses the security concerns in cross-modality person re-identification (ReID) systems, which are more common in practical applications involving infrared cameras. The main challenge in cross-modality ReID is effectively dealing with visual differences between different modalities, such as the grayscale nature of infrared images compared to the color information in visible images. Existing attack methods primarily focus on visible image modality, neglecting the features of other modalities and the variations in data distribution. This study introduces the first exploration into the security of cross-modality ReID models and proposes a universal perturbation attack called Cross-Modality Perturbation Synergy (CMPS). CMPS optimizes perturbations by leveraging gradients from diverse modality data, disrupting the discriminator and reinforcing the differences between modalities. Experiments on the RegDB and SYSU datasets demonstrate the effectiveness of the proposed method, providing insights for future enhancements in the robustness of cross-modality ReID systems. The main contributions include the first investigation into cross-modality ReID security, a novel universal perturbation method, and a cross-modality attack augmentation method using random grayscale transformations. The proposed method shows good transferability across different models and exhibits superior performance compared to existing methods.This paper addresses the security concerns in cross-modality person re-identification (ReID) systems, which are more common in practical applications involving infrared cameras. The main challenge in cross-modality ReID is effectively dealing with visual differences between different modalities, such as the grayscale nature of infrared images compared to the color information in visible images. Existing attack methods primarily focus on visible image modality, neglecting the features of other modalities and the variations in data distribution. This study introduces the first exploration into the security of cross-modality ReID models and proposes a universal perturbation attack called Cross-Modality Perturbation Synergy (CMPS). CMPS optimizes perturbations by leveraging gradients from diverse modality data, disrupting the discriminator and reinforcing the differences between modalities. Experiments on the RegDB and SYSU datasets demonstrate the effectiveness of the proposed method, providing insights for future enhancements in the robustness of cross-modality ReID systems. The main contributions include the first investigation into cross-modality ReID security, a novel universal perturbation method, and a cross-modality attack augmentation method using random grayscale transformations. The proposed method shows good transferability across different models and exhibits superior performance compared to existing methods.
Reach us at info@study.space
[slides] Cross-Modality Perturbation Synergy Attack for Person Re-identification | StudySpace