Protego Masks Facial Signatures to Thwart Retrieval-Based Privacy Intrusions

The increasing prevalence of face recognition technology poses a significant threat to personal privacy, as services now exist that can readily link facial images to a person’s online activity and digital footprint. Ziling Wang, Shuya Yang, and Jialin Lu, from The University of Hong Kong, alongside Ka-Ho Chow and colleagues, address this growing concern with a novel privacy protection method called Protego. This system uniquely encapsulates a user’s facial characteristics into a pose-invariant 2D representation, dynamically transforming it into a realistic 3D mask applied to images before they are shared online. Protego differs from existing approaches by actively increasing the sensitivity of face recognition models, preventing even self-matching of protected images, and experiments demonstrate it substantially reduces retrieval accuracy across various systems, outperforming current methods by at least a factor of two while maintaining exceptional visual quality, particularly in video. Ultimately, this research offers a powerful new tool in the effort to safeguard against the misuse of face recognition for mass surveillance and unwanted identity tracking.

Pose-Invariant Facial Privacy via Adversarial Masks

This research introduces Protego, a new approach to protecting facial privacy by generating adversarial masks that prevent face recognition systems from identifying individuals. The system tackles a significant problem: widespread face recognition technology poses a substantial threat to privacy, with potential for misuse in surveillance. Protego minimizes a user’s digital footprint by generating pose-invariant adversarial masks, preventing retrieval even when attackers use protected images as queries, and maintaining visual coherence in protected images and videos. Key features include adversarial training, 3D facial modeling, and a diversity loss function that encourages varied features in protected images, preventing them from clustering together. The paper demonstrates that Protego significantly reduces the accuracy of face recognition systems, even when the attacker uses a masked image as the query, representing a novel approach to face privacy and a pose-invariant mask generation technique.

Facial Privacy Through Model Sensitivity Amplification

Protego represents a significant advancement in facial privacy protection, addressing concerns about the misuse of face recognition technology and mass surveillance. The system proactively safeguards facial images before they are shared online, preventing unauthorized retrieval of personal information. Unlike existing methods that simply disguise features, Protego fundamentally alters how face recognition models interpret protected images, making them unreliable even when comparing protected images to each other. The core innovation lies in a technique that amplifies the sensitivity of face recognition models, causing even subtle variations in protected faces to produce drastically different results. This disrupts the ability of face recognition systems to find matches, and Protego demonstrably outperforms existing privacy methods, achieving at least double the protection performance while maintaining high visual coherence.

Facial Privacy via Dynamic 3D Masking

Protego offers a new approach to protecting facial images from unauthorized retrieval, addressing growing privacy concerns surrounding face recognition technology. The method encapsulates a user’s facial signature into a pose-invariant 2D representation, dynamically deforming it into a 3D mask applied to images before they are shared online. This significantly reduces the accuracy of face recognition systems and performs notably better than existing privacy protection methods, maintaining natural-looking protected images and videos. The research demonstrates that Protego effectively prevents retrieval even when privacy-invading systems use protected images as search queries, a realistic threat scenario. The system achieves this by amplifying the sensitivity of face recognition models, making it difficult for them to find matches, even amongst images of the same individual, and strengthens protection, particularly for faces in extreme poses or expressions.