深度网络模型可以从视频步态序列中获取人体步态生物特征并识别人物身份,造成严重的隐私泄露安全威胁.现有方法一般通过对视频画面中的人体进行模糊、变形等处理来保护隐私,这些方法可以在一定程度上改变人体外观,但很难改变人物行走姿态,难以逃避深度网络模型的识别,且这种处理往往伴随着对视频质量的严重破坏,降低了视频的视觉可用性.针对该问题,文章提出一种基于轮廓稀疏对抗的视频步态隐私保护算法,通过对步态识别模型的对抗攻击来计算画面中人体轮廓周围的有效修改位置.与传统方法相比,在具有相同隐私保护能力的情况下,该算法减少了对画面的修改,在隐私安全性和视觉可用性上达到了较好的均衡.该算法在公开步态数据库CASIA-B和OUMVLP上对 4 种步态识别模型进行测试,通过与不同步态隐私保护方法对比,验证了该算法在步态隐私保护上的有效性和可用性.
Deep network models can obtain human gait biometrics from video gait sequences and recognize character identities through feature matching,which threatens human privacy.Privacy protection treatments such as blurring and deformation of the human body in video images can to some extent change the appearance of the human body.Still,it is difficult to change the walking posture of the characters and cannot avoid recognition by deep network models.Moreover,this treatment often accompanies serious damage to video quality,reducing the visual usability of the video.In response to this issue,this article proposed a video gait privacy protection algorithm based on sparse adversarial attack on silhouette,which calculates effective modification positions around human silhouette in the image through adversarial attacks on gait recognition models.Compared with traditional methods,this algorithm reduces the modification of images while maintaining the same privacy protection capabilities.The optimal balance between privacy security and visual availability was obtained.The algorithm is tested on four gait recognition models using the public gait datasets CASIA-B and OUMVLP,and previous gait privacy protection methods are implemented and compared,verifying the effectiveness and availability of this algorithm in gait privacy protection.