Privacy-aware video deanonymization: a configurable pipeline for selective reversal with intelligibility preservation

The emergence of Computer Vision (CV) in self-driving vehicles, smart retail, and safety monitoring provides substantial economic benefits. However, it also presents significant privacy concerns for individuals captured on video. Standard techniques to mitigate privacy threats in video processing require some form of video anonymization, such as using blurring or other obfuscation operations to redact human figures from the input data before processing or storage. Yet, the irreversible nature of redaction operations hinders the use of CV in valuable applications, such as quality assurance, safety, or fraud litigation. This paper presents and validates a novel approach for video anonymization that guarantees the anonymity and security of human subjects appearing in the video while maintaining high levels of scene intelligibility. Our approach is based on per-subject episodic key-derived cryptography to securely store data of individuals and enable selective per-subject deanonymization. In this way we can create video streams wherein only consenting subjects can be re-identified. Moreover, our approach can be integrated into CV processing pipelines and, as we also demonstrate, allow for the interchangeability of the visual anonymization techniques, which deliver various degrees of anonymity according to application needs while maintaining significant intelligibility and preserving the scene integrity.