Reversible gender privacy enhancement via adversarial perturbations

Abstract

The significant advancement in deep learning has made it feasible to extract gender from faces accurately. However, such unauthorized extraction would pose potential threats to individual privacy. Existing protection schemes for gender privacy have exhibited satisfactory performance. Nevertheless, they suffer from gender inference from gender-related attributes and fail to support the recovery of the original image. In this paper, we propose a novel gender privacy protection scheme that aims to enhance gender privacy while supporting reversibility. Firstly, our scheme utilizes continuously optimized adversarial perturbations to prevent gender recognition from unauthorized classifiers. Meanwhile, gender-related attributes are concealed for classifiers, which prevents the inference of gender from these attributes, thereby enhancing gender privacy. Moreover, an identity preservation constraint is added to maintain identity preservation. Secondly, reversibility is supported by a reversible image transformation, allowing the perturbations to be securely removed to losslessly recover the original face when required. Extensive experiments demonstrate the effectiveness of our scheme in gender privacy protection, identity preservation, and reversibility.

Keywords: Adversarial perturbations; Gender privacy; Reversible; Soft biometrics.