Privacy-Preserving Analysis for Remote Video Anomaly Detection in Real Life Environments
This paper proposes a novel approach for privacy-preserving surveillance video streams anomaly detection, i.e., situations implying violence, illegal actions, or situations involving hazards. In particular, this approach adopts a privacy-preserving mechanism based on autoencoder neural networks applied in a differential private manner, exploiting three different types of differential private optimizers. Recorded real-world video streams are segmented into data frames, which are compressed into special codes with autoencoders and differential privacy and transmitted to a central server where they get decoded into an anonymized version of the original data frame that can be analyzed to detect anomalies. The anomaly detection algorithm exploits a supervised learning binary classification methodology of extracted contextual, spatial, and motion data on imbalanced datasets. Anomalies are differentiated into ”soft” and ”hard”, and the anomaly detection score is computed based on a sigmoidal function. The proposed methodology has been validated with a set of experiments on a well-known video anomaly dataset: UCF-CRIME. The experiments we conducted on the testbed demonstrate the capability of the system to correctly identify video anomalies, with a consistent privacy gain demonstrated by the strongly reduced ability to identify people from faces in the reconstructed frames.