| 摘 要: 针对目前人脸隐私保护方法难以兼顾隐私保护和数据可用性的问题,提出了一种基于结构保持的匿名 人脸合成方法(Structure Preserving Anonymous Face Synthesis,SPAS)。首先,设计了SPAS网络用于合成匿名人 脸图像;其次,为确保人脸身份信息的隐私安全,提出了多尺度敏感特征匿名方法;最后,利用全重构自监督方法,结 合多重损失函数训练该模型。此外,通过控制非身份属性输入,使得生成的匿名结果更多样化。通过在CelebA-HQ 和LFW 2个权威人脸数据集上进行实验,所提方法的FID分别为26.13和24.56,人脸检测率分别为99.60%和 99.72%。实验结果表明,所提方法在降低身份泄露风险的同时,有效地提高了数据的可用性. |
| 关键词: 人脸图像;生成对抗网络;多尺度特征;敏感特征匿名;隐私保护 |
|
中图分类号: TP391
文献标识码: A
|
|
| Anonymous Face Synthesis Method Based on Structure Preserving |
|
CHEN Zhoujie, CHEN Chao, KUANG Zhenzhong
|
(School of Computer, Hangzhou Dianzi University, Hangzhou 310019, China)
chen_zhoujie@163.com; 211050090@hdu.edu.cn; zzkuang@hdu.edu.cn
|
| Abstract: A Structure Preserving Anonymous Face Synthesis (SPAS) method is proposed to address the challenge of balancing privacy protection and data usability in current face privacy protection methods. Firstly, an SPAS network is designed for synthesizing anonymous face images. Secondly, to ensure the privacy of facial identity information, a multi-scale sensitive feature anonymization method is proposed. Finally, the model is trained using a combination of full reconstruction self-supervised method and multiple loss functions. Additionally, by controlling non-identity-related attribute inputs, the generated anonymous results are diversified. Experiments conducted on the CelebA-HQ and LFW benchmark facial datasets yield FID of 26.13 and 24.56, and face detection rates of 99.60% and 99.72%, respectively. The experiments demonstrate that the proposed method effectively reduces the risk of identity leakage while significantly improving data usability. |
| Keywords: face image; Generative Adversarial Network; multi-scale features; sensitive feature anonymization; privacy protection |
