Adaptive Transformation for Robust Privacy Protection in Video Surveillance

Privacy is a big concern in current video surveillance systems. Due to privacy issues, many strategic places remain unmonitored leading to security threats. The main problem with existing privacy protection methods is that they assume availability of accurate region of interest (RoI) detectors that can detect and hide the privacy sensitive regions such as faces. However, the current detectors are not fully reliable, leading to breaches in privacy protection. In this paper, we propose a privacy protection method that adopts adaptive data transformation involving the use of selective obfuscation and global operations to provide robust privacy even with unreliable detectors. Further, there are many implicit privacy leakage channels that have not been considered by researchers for privacy protection. We block both implicit and explicit channels of privacy leakage. Experimental results show that the proposed method incurs 38% less distortion of the information needed for surveillance in comparison to earlier methods of global transformation; while still providing near-zero privacy loss. 1. Introduction In order to perform privacy-preserving CCTV monitoring, video data should be transformed in such a way that the information leaking the identity is hidden, but the intended surveillance tasks can be accomplished. The traditional approach of data transformation has been to detect the regions of interest (RoI) in the images (e.g., human faces) and selectively obfuscate them. This approach is an unreliable solution as the RoI detectors may sometimes fail. For example, even if a face detector is able to correctly detect the face in 99 (out of 100) frames, the undetected faces in the remaining frame will reveal the identity of the person in the video and result in his/her privacy loss. In other set of works, global operations have been used for data transformation in which the whole video frame is transformed with same intensity, that is, same amount of blurring or quantization [1]. This approach is more appropriate in the context of data publication, where the published surveillance video is used by researchers for testing their algorithms. In contrast to the data publication scenario, CCTV monitoring scenario has different requirements. In the case of CCTV monitoring, a human operator is required to watch the surveillance video feeds; although automated techniques may run in the background as shown in Figure 1. The automatic analysis can be performed using the original data, which is not accessible for viewing, unlike data publication. The original data may be