Abstract
Person re-identification (ReID) is to associate images of individuals from different camera views against cross-view variations. Like other surveillance technologies, Re-ID faces serious privacy challenges, particularly the potential for unauthorized tracking. Although various tasks (e.g., face recognition) have developed machine unlearning techniques to address privacy concerns, such methods have not yet been explored within the Re-ID field. In this work, we pioneer the exploration of the person de-reidentification (De-ReID) problem and present its inherent challenges. In the context of ReID, De-ReID is to unlearn the knowledge about accurately matching specific persons so that these "unlearned persons" cannot be re-identified across cameras for privacy guarantee. The primary challenge is to achieve the unlearning without degrading the identity-discriminative feature embeddings to ensure the model's utility. To address this, we formulate a De-ReID framework that utilizes a labeled dataset of unlearned persons for unlearning and an unlabeled dataset of accessible persons for knowledge preservation. Instead of unlearning based on (pseudo) identity labels, we introduce a variation-guided identity shift mechanism that unlearns the specific persons by fitting the variations in their images while preserving ReID ability on other persons by overcoming the variations in images of accessible persons. As a result, the model shifts the unlearned persons to a feature space that is vulnerable to cross-view variations. Extensive experiments on benchmarks demonstrate the superiority of our method.
Introduction
Person Re-identification (ReID) is a cornerstone of modern surveillance, allowing systems to track individuals across networks of non-overlapping camera views. While useful for security, it poses significant privacy risks and ethical concerns, often clashing with regulations like the EU’s GDPR. The challenge is to stop a model from tracking a specific person (e.g., a resident in a monitored complex) without breaking the model’s ability to track others.
Existing machine unlearning techniques mostly focus on classification tasks. However, ReID relies on feature embeddings—mapping images to a high-dimensional space where images of the same person are close together. Simply "forgetting" a label is insufficient; one must fundamentally alter how the model perceives specific features across different environments (lighting, angles, etc.) without collapsing the entire feature space.
Methodology
The proposed framework treats unlearning as an Identity Shift. Instead of erasing data, the model's sensitivity to image variations is manipulated to create a bifurcated feature space.
Variation-guided Identity Shift (VIS)
The core idea is based on the nature of ReID: a good model is "view-invariant." To unlearn someone, the authors make the model view-variant for that person:
- For Unlearned Persons: The model is trained to fit variations. It is guided to distinguish between an original image and its augmented version (different brightness, color, etc.). This pushes the person's features out of the stable, view-invariant space.
- For Accessible Persons: The model is trained to overcome variations, pulling original and augmented images closer together to preserve standard ReID performance.
Relation Regularization
To ensure the model doesn't lose general utility, two additional constraints are added:
- Triplet Constraint ($L_{TRC}$): Ensures images of the same unlearned person are pushed further apart than they are from images of different (accessible) people.
- Relation Consistent Regularization ($L_{RCR}$): Acts as a "knowledge anchor," forcing the model to maintain the same relative similarity scores for accessible people as the original, pre-unlearned model.
Experimental Results
The authors evaluated their method on standard ReID benchmarks (MSMT17 and Market-1501) using H-Mean (the harmonic mean of unlearning effectiveness and ReID maintenance) as the primary metric.
Key findings include:
- Superiority over Baselines: On MSMT17 (with 100 unlearned persons), the method achieved an H-Mean of 66.7%, vastly outperforming general unlearning methods like GS-LoRA (28.3%).
- Effective Forgetting: Rank-1 accuracy for unlearned persons dropped to near-zero levels (e.g., 4.6% in MSMT17), effectively "blinding" the model to those individuals.
- Utility Preservation: The model maintained high Rank-1 accuracy for accessible persons (e.g., 77.0% on MSMT17), proving that privacy can be achieved without sacrificing total system utility.
Conclusion
This work introduces De-ReID as a novel approach to privacy in surveillance. By treating unlearning as an image-specificity problem rather than a label-erasure problem, the Variation-guided Identity Shift successfully isolates target identities in the feature space. Future work will explore unsupervised or zero-shot settings where no images of the unlearned person are available during training.
Summary updated