.png){kind=icon}
.png){kind=icon}
Return
Revisiting the machine unlearning ecosystem in vision: From afterthought to a lifecycle perspective
.png){kind=icon}
PRE
AI
E
M
B
T
A
DOI：10.1016/j.imavis.2026.106059.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Vision systems such as biometric recognition, medical imaging platforms, surveillance cameras, and autonomous driving perception systems can encode sensitive identity attributes and contextual information within high-dimensional representations, creating privacy and compliance risks. As these systems operate in dynamic, real-world environments, requirements such as privacy preservation and regulatory compliance become increasingly critical. Machine Unlearning (MU) offers a promising paradigm for reducing the influence of selected information in trained models while preserving acceptable performance. Focusing on security-driven vision application domains, MU can support (1) removing the influence of identity-specific knowledge (i.e., through model weight updates) from biometric systems, (2) addressing copyright and Intellectual Property (IP) challenges arising from the use of proprietary visual data, and (3) reducing the influence of sensitive information that vision models may implicitly encode from visual inputs, such as Personally Identifiable Information (PII), biometric identifiers, and contextual cues present in images or video data. However, there is a fundamental discrepancy between modern vision systems, which are constantly evolving over time (e.g., incremental learning or updates), iterative, and deployment-driven rather than static pipelines, and MU, which is still largely treated as an isolated algorithmic primitive. This paper argues that MU must be viewed as a relevant component of the vision AI lifecycle; although the methodologies required to achieve this integration remain an open challenge for the research community.
Journal
IF:
4.2
Papers:
4.0K
Citations:
6.7K
Organization
Cited Papers