Password-conditioned Anonymization and Deanonymization with Face Identity Transformers (1911.11759v4)

Abstract: Cameras are prevalent in our daily lives, and enable many useful systems built upon computer vision technologies such as smart cameras and home robots for service applications. However, there is also an increasing societal concern as the captured images/videos may contain privacy-sensitive information (e.g., face identity). We propose a novel face identity transformer which enables automated photo-realistic password-based anonymization as well as deanonymization of human faces appearing in visual data. Our face identity transformer is trained to (1) remove face identity information after anonymization, (2) make the recovery of the original face possible when given the correct password, and (3) return a wrong--but photo-realistic--face given a wrong password. Extensive experiments show that our approach enables multimodal password-conditioned face anonymizations and deanonymizations, without sacrificing privacy compared to existing anonymization approaches.

• The paper introduces a transformer-based face identity model that uses password conditioning for dynamic anonymization and deanonymization.
• It employs a multi-task learning strategy with GANs to generate realistic, variable face outputs based on correct or incorrect passwords.
• Empirical evaluations on datasets like CASIA, LFW, and FFHQ confirm that the model preserves privacy while enabling secure identity recovery.

Overview of Password-conditioned Anonymization and Deanonymization with Face Identity Transformers

This paper introduces an innovative approach to face anonymization and deanonymization using a password-conditioned face identity transformer. The proposed system effectively balances individual privacy with the utility of being able to recover the original identity under controlled conditions. The authors present a transformer model that leverages discrete passwords to conditionally modify face identities within visual data, thereby ensuring privacy preservation while maintaining the ability to revert to the genuine identity when necessary.

Methodological Framework

The central contribution of this paper is the face identity transformer, which is capable of both anonymizing and deanonymizing human faces in a photo-realistic manner driven by password conditioning. Unlike traditional anonymization methods such as downsampling or pixel masking, this approach retains the usability of the visual data by creating realistic face representations devoid of the actual identity.

The transformer model operates in three key steps:

1. Anonymization: Removes identifiable information from the face image by altering the identity according to a password.
2. Deanonymization: Restores the original image if the correct recovery password is provided.
3. Incorrect Deanonymization: Generates a third, yet still realistic, face when an incorrect recovery password is supplied. This is integral for maintaining security even in the presence of potential password guessing attacks.

Technical Implementation

The authors employ a multi-task learning strategy that includes a face classification adversarial loss and a feature dissimilarity component. This combination ensures that the anonymized faces differ significantly across different passwords, thereby preventing adversaries from easily reverse-engineering the transformations. The use of a Generative Adversarial Network (GAN) framework further aids in maintaining photo-realism in the anonymized outputs.

An additional noteworthy aspect is the identification of a password format and embedding scheme that ensures diverse identity outputs. This factor enables the model to conditionally alter faces to produce a broad range of plausible identities. Moreover, by storing only the anonymized faces, the method inherently increases data security by not retaining the original images on disk.

Experimental Evaluation

The proposed method is benchmarked against several existing anonymization techniques across datasets such as CASIA, LFW, and FFHQ. In terms of face verification, the model presents comparable anonymization and superior deanonymization capabilities. The paper's empirical results demonstrate effective multimodal face manipulation, implying that the transformer can reliably generate varied and realistic identities. Moreover, human studies via Amazon Mechanical Turk (AMT) assess the perceptual realism of the generated faces, corroborating the method's photo-realistic claims.

Implications and Speculative Outlook

This research introduces a practical solution for privacy-preserving image processing, particularly relevant in contexts like surveillance, social media, or any scenario involving sensitive visual data. The ability to conditionally reverse anonymization opens avenues for controlled scenarios where identity retrieval is essential, such as law enforcement inquiries or familial connections.

Future developments might expand upon this approach, potentially incorporating more robust adversarial defenses ensuring security against evolving threats. Furthermore, improving the transformer model to enhance consistency across a broader set of image resolutions and conditions can significantly boost its applicability in real-world settings.

In conclusion, this paper pioneers a sophisticated balance between privacy and accessibility using a password-dependent model for face anonymization and deanonymization, providing significant advances in the field of privacy-preserving computer vision technology.