基于对抗学习与协同优化的身份匿名方法  

Adversarial Learning and Collaborative Optimization for Identity Anonymization

作　　者：刘玉彤 Liu Yutong(Heilongjiang University of Industry and Business,Harbin 150025,China)

机构地区：[1]黑龙江工商学院,哈尔滨150025

出　　处：《黑龙江科学》2024年第4期109-112,共4页Heilongjiang Science

摘　　要：目前,深度学习方法通常依赖于大规模的数据集,这些数据集不可避免地涉及个人隐私,从而引起隐私泄露问题。为解决以上问题,身份匿名方法应运而生。匿名方法使用人脸检测模型掩盖原始人脸,对人脸区域进行重新生成,但额外的人脸检测模型显著增加了训练成本及模型推理时间。针对此问题,提出了一种基于对抗学习与协同优化的身份匿名方法。该方法在生成对抗网络中引入识别损失,匿名原始图像的身份,并增加重建损失以保持原始图像的背景,通过平衡以上两个损失,可在保持背景不变的前提下对身份进行匿名。将此方法在CelebA数据集上进行大量测试,实验结果表明,相比现有方法,基于对抗学习与协同优化的身份匿名方法在生成质量、匿名效果及推理速度等方面表现出显著的优越性,不仅优于现有的一对一方法,还优于多对一方法。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 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]