Face De-Identification: A Domain-Centric Survey from Capture to Processing

Face de-identification operates across three distinct stages — physical (wearables and materials), sensor (privacy built into cameras), and digital (post-processing) — with the survey noting that recent advances integrate privacy protection directly into image capture rather than relying solely on after-the-fact modifications, offering earlier intervention points in the data pipeline.

The field lacks standardized evaluation benchmarks, which the authors identify as a "critical need" — this gap means different de-identification methods are currently tested using inconsistent protocols, making it difficult to compare which approaches actually work best or to ensure privacy protections are genuinely effective across different scenarios.

De-identification must balance privacy with utility, as the goal is to remove identifying facial features while still preserving enough image quality for downstream tasks like medical diagnosis or security monitoring — this tension means overly aggressive privacy protection can render data useless, while insufficient protection leaves people identifiable.

Physical and sensor-level privacy approaches represent an emerging frontier beyond traditional digital image editing, suggesting that responsible AI practitioners should consider privacy protection as a system-wide concern starting from hardware design rather than treating it as a software problem to solve after data collection.

The rapid evolution of face de-identification reflects growing recognition that privacy and responsible AI require proactive technical solutions, with the survey documenting how the field has expanded from simple pixel-modification techniques to sophisticated approaches spanning multiple domains of the data acquisition pipeline.