Abstract
Advancements in deep learning have improved anomaly detection in surveillancevideos, yet they raise urgent privacy concerns due to the collection ofsensitive human data. In this paper, we present a comprehensive analysis ofanomaly detection performance under four human anonymization techniques,including blurring, masking, encryption, and avatar replacement, applied to theUCF-Crime dataset. We evaluate four anomaly detection methods, MGFN, UR-DMU,BN-WVAD, and PEL4VAD, on the anonymized UCF-Crime to reveal how each methodresponds to different obfuscation techniques. Experimental results demonstratethat anomaly detection remains viable under anonymized data and is dependent onthe algorithmic design and the learning strategy. For instance, under certainanonymization patterns, such as encryption and masking, some modelsinadvertently achieve higher AUC performance compared to raw data, due to thestrong responsiveness of their algorithmic components to these noise patterns.These results highlight the algorithm-specific sensitivities to anonymizationand emphasize the trade-off between preserving privacy and maintainingdetection utility. Furthermore, we compare these conventional anonymizationtechniques with the emerging privacy-by-design solutions, highlighting an oftenoverlooked trade-off between robust privacy protection and utility flexibility.Through comprehensive experiments and analyses, this study provides acompelling benchmark and insights into balancing human privacy with the demandsof anomaly detection.