notesum.ai

Published at November 4

Pseudo-Probability Unlearning: Towards Efficient and Privacy-Preserving Machine Unlearning

cs.LG

cs.AI

Released Date: November 4, 2024

Authors: Zihao Zhao¹, Yijiang Li², Yuchen Yang¹, Wenqing Zhang¹, Nuno Vasconcelos², Yinzhi Cao¹

Aff.: ¹Johns Hopkins University; ²University of California San Diego

Arxiv: http://arxiv.org/abs/2411.02622v1

$Refer to caption$

	CIFAR-5			Lacuna-5
Model	Test error ( $\downarrow$ )	Retain error ( $\downarrow$ )	Forget error ( $\uparrow$ )	Test error ( $\downarrow$ )	Retain error ( $\downarrow$ )	Forget error ( $\uparrow$ )
Retrain	24.90	$\mathbf{0.00}$	28.80	5.80	$\mathbf{0.00}$	4.80
Original	24.20	$\mathbf{0.00}$	0.00	5.70	$\mathbf{0.00}$	0.00
Finetune	24.30	$\mathbf{0.00}$	0.00	5.60	$\mathbf{0.00}$	0.00
Fisher	31.60	14.00	4.80	6.70	14.00	6.40
NTK	24.40	$\mathbf{0.00}$	22.40	5.60	$\mathbf{0.00}$	0.00
NegGrad+	25.50	$\mathbf{0.00}$	41.3	6.10	$\mathbf{0.00}$	1.30
CF-k	22.60	$\mathbf{0.00}$	0.00	5.80	$\mathbf{0.00}$	0.00
EU-k	23.50	$\mathbf{0.00}$	10.70	5.90	$\mathbf{0.00}$	0.00
Bad-T	22.73	5.12	8.00	5.00	8.64	0.14
SCRUB	24.20	$\mathbf{0.00}$	40.80	6.20	$\mathbf{0.00}$	24.80
PPU w/ random	$\mathbf{22.00}$	$\mathbf{0.00}$	$\mathbf{80.00}$	$\mathbf{2.20}$	$\mathbf{0.00}$	64.00
PPU w/ uniform	27.00	0.21	$\mathbf{80.00}$	2.80	0.42	$\mathbf{68.00}$