基于对抗性注意力机制的行人再识别攻击方法
Pedestrian re-identification attack based on adversarial attention mechanism
张琳钰 1王进 1万杰 1刘国庆2
作者信息
- 1. 南通大学信息科学技术学院,江苏南通 226000
- 2. 中天智能装备有限公司 研发部,江苏 南通 226010
- 折叠
摘要
针对行人再识别系统鲁棒性的问题,提出一种基于对抗性注意力机制的行人再识别攻击方法.将通道注意力机制嵌入改造过的ResNet50网络,提取整个行人的显著特征;利用重新设计的掩码打破原注意力权重,混淆关注区域,降低行人再识别模型的识别能力;联合对抗三元组损失和对抗ID损失监督网络训练,得到行人图像匹配对错误排序,更适用行人再识别中对抗攻击的排序问题.在Market-1501和DukeMTMC-reID数据集上模拟攻击,实验结果表明,该方法能将主流行人再识别系统的Rank-1和平均精度均值(mAP)降低至0.1％,为设计行人再识别防御方法提高鲁棒性提供新启示.
Abstract
Aiming at the problem of the robustness of pedestrian re-identification system,a pedestrian re-identification attack method based on adversarial attention mechanism was presented.The channel attention mechanism was embedded in the modified ResNet50 network to extract the salient features of the entire pedestrian.The redesigned mask was used to break the original attention weight and confuse the attention area,thus the recognition ability of the pedestrian re-identification model was reduced.Training of the network was supervised by the combination of adversarial triplet loss and adversarial ID loss,the wrong ranking of pedestrian image matching pairs was obtained,which was more suitable for the ranking problem of adversarial attack in pedes-trian re-identification.Results of the simulation on Market-1501 and DukeMTMC-reID datasets show that Rank-1 and mean average precision(mAP)of the mainstream pedestrian re-identification system are reduced to 0.1％,which is a new enlighten-ment to design person re-identification defense to improve the robustness.
关键词
行人再识别/对抗攻击/掩码/注意力机制/注意力权重/排序损失/鲁棒性Key words
pedestrian re-identification/adversarial attack/mask/attention mechanism/attentional weight/ranking loss/ro-bustness引用本文复制引用
基金项目
国家自然科学基金青年基金(62002179)
南通市基础研究计划(2022)(JC22022061)
出版年
2024
NETL
NSTL
维普
万方数据