notesum.ai

Published at November 25

Can Encrypted Images Still Train Neural Networks? Investigating Image Information and Random Vortex Transformation

cs.CR

Released Date: November 25, 2024

Authors: XiaoKai Cao, WenJin Mo, ChangDong Wang, JianHuang Lai, Qiong Huang

Arxiv: http://arxiv.org/abs/2411.16207v1

Refer to caption

Model	MNIST			Fashion			CIFAR-10
Model	Origin	Vortex	Random	Origin	Vortex	Random	Origin	Vortex	Random
ResNet-18	99.20%	$98.89\%^{{\color[rgb]{0,1,0}\downarrow 0.31\%}}$	$28.99\%^{{\color[rgb]{0,1,0}\downarrow 70.21\%}}$	94.12%	$92.85\%^{{\color[rgb]{0,1,0}\downarrow 1.27\%}}$	$40.10\%^{{\color[rgb]{0,1,0}\downarrow 54.02\%}}$	92.80%	$87.56\%^{{\color[rgb]{0,1,0}\downarrow 5.24\%}}$	$42.80\%^{{\color[rgb]{0,1,0}\downarrow 50.0\%}}$
ViT	99.47%	$99.15\%^{{\color[rgb]{0,1,0}\downarrow 0.32\%}}$	$31.84\%^{{\color[rgb]{0,1,0}\downarrow 67.63\%}}$	94.95%	$92.61\%^{{\color[rgb]{0,1,0}\downarrow 2.34\%}}$	$56.19\%^{{\color[rgb]{0,1,0}\downarrow 38.76\%}}$	98.50%	$92.04\%^{{\color[rgb]{0,1,0}\downarrow 6.46\%}}$	$50.07\%^{{\color[rgb]{0,1,0}\downarrow 48.43\%}}$
$\Upsilon(\cdot)$	100.00%	$95.25\%^{{\color[rgb]{0,1,0}\downarrow 4.75\%}}$	$67.34\%^{{\color[rgb]{0,1,0}\downarrow 32.66\%}}$	100.00%	$90.04\%^{{\color[rgb]{0,1,0}\downarrow 9.96\%}}$	$67.23\%^{{\color[rgb]{0,1,0}\downarrow 32.77\%}}$	100.00%	$97.38\%^{{\color[rgb]{0,1,0}\downarrow 2.62\%}}$	$67.08\%^{{\color[rgb]{0,1,0}\downarrow 32.92\%}}$