notesum.ai

Published at October 30

CLIPErase: Efficient Unlearning of Visual-Textual Associations in CLIP

cs.CV

cs.AI

cs.LG

Released Date: October 30, 2024

Authors: Tianyu Yang¹, Lisen Dai², Zheyuan Liu¹, Xiangqi Wang¹, Meng Jiang¹, Yapeng Tian³, Xiangliang Zhang¹

Aff.: ¹University of Notre Dame; ²Columbia University; ³The University of Texas at Dallas

Arxiv: http://arxiv.org/abs/2410.23330v1

Refer to caption

Method	Zero-shot Prediction		Zero-shot Retrieval
Method	Acc. $D_{f}$ $\downarrow$	Acc. $D_{r}$ $\uparrow$	Acc. $D_{f}$ $\downarrow$	Acc. $D_{r}$ $\uparrow$
CLIP	86.08	72.85	88.61	73.43
CLIP+GA	11.81	66.26	5.35	67.31
CLIP+GradDiff	7.31	70.94	9.89	71.79
CLIP+KL	2.19	61.75	5.74	65.87
CLIP+ENMN	0.00	12.46	0.00	17.94
CLIPErase	0.00	89.41	0.00	91.38