Currently,deep learning methods often rely on large-scale datasets,which inevitably involve individual privacy concerns,leading to privacy leakage issues.In response to the challenges,identity anonymization techniques have emerged.Existing anonymization approaches initially employ facial detection models to conceal the original faces,followed by regeneration of the facial regions.While these methods have demonstrated promising performance,the additional facial detection models significantly escalate training costs and inference time.To mitigate the aforementioned challenges,the study introduces an identity anonymization framework based on adversarial learning and collaborative optimization.The proposed method incorporates an identification loss within the generative adversarial network to anonymize the identity of the original images,while augmenting a reconstruction loss to preserve the background of the original images.By balancing these two losses,the proposed method achieves identity anonymization while maintaining background consistency.The proposed method underwent extensive testing on the CelebA dataset.Experimental results demonstrate that,in comparison to existing methods,the proposed method based on adversarial learning and collaborative optimization exhibits significant superiority in terms of generation quality,anonymization effectiveness,and inference speed.It not only surpasses existing one-to-one methods but also outperforms many-to-one methods.
关键词
深度学习/生成对抗网络/图像生成/身份匿名化
Key words
Deep Learning/Generative adversarial networks/Image generation/Identity anonymization
维普
万方数据