Targeted Forgetting of Image Subgroups in CLIP Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Zeliang, Liu, Gaowen, Fleming, Charles, Kompella, Ramana Rao, Xu, Chenliang
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909635243409408
author Zhang, Zeliang
Liu, Gaowen
Fleming, Charles
Kompella, Ramana Rao
Xu, Chenliang
author_facet Zhang, Zeliang
Liu, Gaowen
Fleming, Charles
Kompella, Ramana Rao
Xu, Chenliang
contents Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class, without access to pre-trained data, while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting, retaining samples, and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting.
format Preprint
id arxiv_https___arxiv_org_abs_2506_03117
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Targeted Forgetting of Image Subgroups in CLIP Models
Zhang, Zeliang
Liu, Gaowen
Fleming, Charles
Kompella, Ramana Rao
Xu, Chenliang
Computer Vision and Pattern Recognition
Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class, without access to pre-trained data, while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting, retaining samples, and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting.
title Targeted Forgetting of Image Subgroups in CLIP Models
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2506.03117