.png){kind=icon}
Return
Vec2Null: Structured Identity Vectorization and Nulling for One-Shot Face Unlearning
.png){kind=icon}
PRE
AI
DOI：10.1109/tifs.2026.3734767.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Machine unlearning has emerged as a critical mechanism for personal privacy compliance. However, most existing methods require access to the original training dataset, which is impractical under stringent data regulations. To address this, we investigate the practical problem of one-shot personal identity unlearning, where an unlearning request is fulfilled using exclusively a new face image provided by the user. In this regime, existing methods are ill-posed: optimizing directly on a single, noisy support view fails to generalize across the target identity’s manifold, inevitably resulting in either incomplete erasure or utility degradation. To bridge this single-view to manifold gap, we propose Vec2Null, a geometry-guided framework for one-shot identity unlearning on face-analytics models. Instead of unlearning directly from the support image, Vec2Null first vectorizes the single-shot support image into a stable identity-level direction and then performs a retain-safe nulling update. We first introduce an Unlearning-Oriented Face Representation Network (UnoFace) optimized via a novel asymmetric One-Shot Proxy objective. UnoFace maps arbitrary single views to stable, low-variance proxy directions for the target identity manifold. Building upon this, Vec2Null employs a context-aware meta-learner to synthesize robust parameter updates by fusing the target model’s internal response with the stable unlearning directions provided by UnoFace. Finally, the unlearning updates are applied via a projection operator to map the update away from retain-sensitive subspaces. Extensive experiments across multiple benchmarks demonstrate that Vec2Null achieves state-of-the-art performance in one-shot face unlearning. Code will be released at: github.com/CV-AC/Vec2Null.
Keywords:
Machine unlearning
one-shot face unlearning
privacy protection
identity vectorization
Journal
IF:
8
Papers:
5.3K
Citations:
2.3W
Organization
Cited Papers
No cited papers available