Wednesday 09 April 2025
Scientists have made a significant breakthrough in developing a new framework for generating customized portraits that can deceive malicious face recognition systems. This innovative approach, called Adv-CPG, combines facial privacy protection and fine-grained portrait generation to create highly realistic images.
The traditional method of generating customized portraits involves using diffusion models to synthesize an image based on a textual prompt. However, these models often lack the ability to protect the subject’s identity from being tracked or misused by malicious actors. To address this issue, researchers have incorporated facial privacy protection modules into the portrait generation process.
Adv-CPG uses a lightweight local ID encryptor and encryption enhancer to progressively double-layer encrypt the target identity, ensuring that the generated portrait does not reveal sensitive information about the subject. This approach not only protects the subject’s identity but also enables fine-grained control over the facial features, allowing for highly realistic portraits.
The Adv-CPG framework consists of two stages: progressive facial privacy protection and fine-grained customized portrait generation. In the first stage, the original text prompt is used to introduce contextual information about the subject, while the ID encryptor and encryption enhancer work together to protect the target identity. The second stage utilizes the augmented text prompt to generate a highly realistic portrait that adheres to the textual description.
The researchers tested Adv-CPG on various facial recognition systems and found that it achieved optimal or suboptimal results in protecting the subject’s identity. They also demonstrated the framework’s robustness by using diverse FR models, including closed-source systems like Face++ and Aliyun.
One of the key advantages of Adv-CPG is its ability to adapt to different diffusion models. By training the ID projector in the encryption enhancer on a specific diffusion model, the framework can seamlessly integrate with various T2I models. This flexibility makes Adv-CPG a versatile tool for generating customized portraits that meet diverse requirements.
The Adv-CPG framework has significant implications for facial privacy protection and portrait generation. It provides a new approach to protecting sensitive information about individuals while generating highly realistic portraits that adhere to textual descriptions. As the use of facial recognition systems continues to grow, Adv-CPG offers a promising solution for ensuring the privacy and security of individuals in these systems.
The researchers plan to further develop and refine the Adv-CPG framework by exploring new diffusion models and facial recognition systems. They also intend to investigate applications of Adv-CPG in various domains, such as art generation and entertainment.
Cite this article: “Unlocking Private Portraits: A Novel Framework for Adversarial Facial Privacy Protection”, The Science Archive, 2025.
Portraits, Facial Recognition, Privacy Protection, Portrait Generation, Diffusion Models, Facial Features, Encryption, Identity Protection, Customized Portraits, T2I Models
