Toward Generating Unlearnable Examples for Open-Set Face Recognition

초록

Constructing massive face recognition (FR) datasets without users' consent raises concerns about the responsible collection and use of facial data. To prevent such unauthorized exploitation, we propose the first unlearnable example (UE) generator for open-set FR. It injects imperceptible noise into faces, inducing shortcuts that prevent the FR model from learning semantic features from protected data. While UEs have been extensively studied in image classification, existing techniques, e.g., class-bound UEs, do not apply to open-set FR: train and test identities are disjoint, and the model learns embedding geometry rather than class boundaries. We address this gap by proposing three losses: feature-level smudging loss, residual alignment loss with face foundation models, and augmentation consistency loss. With these losses, the resulting UEs directly induce shortcuts in the FR model's embedding space, while generalizing to unseen faces across datasets. Extensive experiments demonstrate the effectiveness of our method on multiple FR models, as well as the imperceptibility of the injected noise.

키워드