Protecting Feature Privacy in Person Re-identification | JoVE Visualize

Area of Science:

Computer Vision
Artificial Intelligence
Cybersecurity

Background:

Deep learning methods dominate person re-identification (ReID), extracting features for efficient querying.
Current ReID feature extraction methods fail to adequately protect privacy, as features can be reversed to reconstruct original images.
Existing privacy-preserving methods struggle to properly assess private information due to reconstruction distribution mismatches.

Purpose of the Study:

To develop a novel Generative Adversarial Network (GAN)-based person ReID model that enhances feature privacy against reversal attacks.
To maintain the utility of person ReID while effectively protecting sensitive information within extracted features.
To introduce a new metric, the utility-reversibility ratio (URR), for evaluating the balance between ReID performance and privacy protection.

Main Methods:

Implemented a GAN-based feature reversal module integrated with a conventional ReID feature extraction module.
Developed a two-step training and lazy update strategy to stabilize the training of the dual adversarial objectives.
Utilized a novel metric, the utility-reversibility ratio (URR), to quantify the model's performance in balancing privacy and utility.

Main Results:

The proposed GAN-based ReID model effectively protects feature privacy against reversal attacks with minimal impact on ReID accuracy.
The two-step training and lazy update strategy successfully stabilized the optimization process for the ReID and privacy protection objectives.
Extensive experiments validated the model's superior performance in balancing privacy protection and ReID utility compared to existing methods, including effectiveness with diffusion models.

Conclusions:

The novel GAN-based person ReID model offers a robust solution for privacy-preserving feature extraction.
The developed training strategy and URR metric provide effective means to address the challenges of optimizing for both ReID utility and privacy.
The model demonstrates significant potential for real-world applications requiring secure and accurate person re-identification.