作者：Tao Wang, Xiaoyu Chen  Zhiquan Liu,Shihong Yao

出版刊物：Information Fusion

出版时间：2025年

内容摘要：

GenAI harnesses vast amounts of personal data, specifically facial images, to achieve remarkable generative capabilities. As it continues to evolve and integrate into various applications, this practice has raised significant privacy concerns. Current state-of-the-art facial privacy-preserving methods predominantly utilize a generative paradigm, where sensitive facial features, such as identity information or specific attributes, are extracted and subsequently sanitized. The fusion of these sanitized feature with other non- or soft-biometric attributes is executed to generate privacy-ensured synthetic faces. However, these methods always generate faces of insufficient security and low realism. To address these challenges, we propose DiffuShield, a novel facial privacy-preserving method that integrates identity and attribute hiding based on diffusion model. DiffuShield enables configurable, confidential, and practical facial privacy preservation by allowing selective preservation of identities and sensitive attributes. It adopts the diffusion model as the backbone generative network and introduces identity and attribute encoders as conditional inputs. These two encoders can efficiently decompose facial representations while incorporating differential privacy mechanisms by introducing Laplacian noise, thereby achieving flexible facial privacy preservation. Our experimental results demonstrate that DiffuShield not only excels in preserving privacy but also maintains image quality and compatibility with various computer vision tasks. This balance of privacy and performance positions DiffuShield as a robust solution for real-world applications requiring both privacy preservation and functionality.

生成式人工智能（GenAI）通过处理海量个人数据（特别是面部图像）来获得强大的图像生成能力。随着该技术不断演进并广泛应用于各领域，这种数据使用方式引发了日益严峻的隐私泄露隐患。目前主流的面部隐私保护方案主要基于生成式框架：先提取面部特征中的敏感信息（如身份标识或特定属性），经过脱敏处理后，再与其他非生物特征进行融合，最终生成具有隐私保护效果的合成人脸。但现有方法生成的人脸图像往往存在两大缺陷：安全防护强度不足，且视觉真实感较差。

为解决这些问题，我们创新性地提出DiffuShield方案——基于扩散模型的面部隐私保护技术，能够同步实现身份信息与敏感属性的双重隐匿。该方案支持选择性保留部分身份特征与敏感属性，使隐私保护具备灵活配置、安全可靠、便于落地三大特性。具体而言，DiffuShield以扩散模型为核心生成架构，引入身份编码器与属性编码器作为条件输入模块。通过注入拉普拉斯噪声实现差分隐私保护，两个编码器可高效解构面部特征，从而实现精准可控的隐私保护。实验数据表明，DiffuShield在实现高效隐私保护的同时，不仅能保持优异的图像质量，还能兼容各类计算机视觉任务的应用需求。这种在隐私安全与功能效能之间取得的平衡，使DiffuShield成为兼顾隐私保护与实用价值的创新解决方案。