Lightweight Face Image Anonymization via Differential Privacy and Feature Decomposition

Advances in deep learning have driven the proliferation of personalized image services: users upload face images to cloud servers running deep models to obtain treatments such as makeup transfer and wrinkle removal. Growing concern for personal privacy, however, has raised fears of identity leakage. Image-encoding perturbation methods based on generative adversarial networks (GANs) can obfuscate facial identity and protect identity privacy. Existing methods typically do not separate identity features from appearance features within facial encodings; approaches that rely on multiple feature-extraction networks to preserve appearance incur high computational cost and are difficult to deploy locally, while perturbation schemes based on random rotations of feature vectors offer limited protection. To address these issues, we propose a face anonymization and identity-reconstruction framework and design DPDF-GAN, a face image anonymization model that integrates differential privacy with feature-space decomposition. Our method decomposes facial encodings into identity and appearance components, applies differential-privacy perturbations to the identity features, and generates anonymized faces from the perturbed encodings. We also develop an identity reconstruction procedure that reassembles a new encoding from locally stored original identity features and cloud-processed appearance features, and uses the DPDF-GAN decoder to reconstruct the user’s face image. Theoretical analysis and experiments demonstrate that the proposed scheme preserves facial appearance while protecting identity privacy, reduces computational cost, and supports lightweight client-side deployment.