A Semantic-Convolutional Generative Adversarial Network for privacy-preserving image de-identification (SCP-GAN)

To address the long-standing tension in image de-identification between privacy strength and downstream utility, we propose a Semantic-Convolutional Privacy-preserving Generative Adversarial Network (SCP-GAN). SCP-GAN explicitly disentangles identity-related cues from semantic content using a convolutional feature extractor, and further introduces an identity-confusion branch to impose controllable perturbations on identity representations directly in the feature space. Coupled with a semantics-guided generator, the proposed framework synthesizes high-fidelity de-identified images that effectively suppress identity leakage while preserving semantic structure and visual realism. Experiments on two public benchmarks, CelebA and UTKFace, demonstrate that SCP-GAN consistently outperforms several recent representative baselines in terms of resistance to identity recognition attacks, semantic attribute retention, and generation quality. Moreover, SCP-GAN achieves an overall improvement of approximately 15–20% in the holistic trade-off between privacy protection and image utility compared with mainstream competing methods. Ablation studies further confirm the indispensable contributions of individual modules to the joint optimization of privacy preservation and generative quality. Collectively, SCP-GAN offers a structured and controllable solution for privacy-preserving image generation with strong generalization capability.