Obscrowd: Joint Face-Localized Identity Protection for Multi-Portrait Images under Generative Editing

Generative image editors can modify group photographs while preserving biometric cues that allow face-recognition systems to identify the depicted individuals. We present Obscrowd, a method that jointly optimizes a single perturbation localized to all detected faces through soft masks. A diffusion teacher-student pipeline with inference-time refinement optimizes identity disruption across simulated edits using an ensemble of face recognizers. We evaluate Obscrowd on group photographs and single portraits, comparing protected faces with matched unprotected controls to distinguish protection from identity loss caused by editing itself. Evaluation examines both crop-edit orderings, transfer to a held-out recognizer, and the effects of perturbation budget and perceptual cost. Across the evaluated editing and processing conditions, Obscrowd reduces biometric similarity significantly relative to matched unprotected controls. Under crop-then-edit evaluation with an instruction-based editor, mean cosine similarity on group photographs falls from 0.662 to 0.287. Joint optimization also achieves protection comparable to independent per-face optimization using approximately one-quarter of the compute. Alongside these gains, protection is limited to detected faces and weakens as wider crops reduce the protected face’s share of the recognizer input. These findings support joint optimization as a computationally efficient approach to protecting multiple faces under generative editing, while identifying cropping and adaptive attacks as priorities for further investigation.