notesum.ai

Published at December 9

Take Fake as Real: Realistic-like Robust Black-box Adversarial Attack to Evade AIGC Detection

cs.CV

Released Date: December 9, 2024

Authors: Caiyun Xie¹, Dengpan Ye, Yunming Zhang, Long Tang, Yunna Lv, Jiacheng Deng, Jiawei Song

Aff.: ¹University

Arxiv: http://arxiv.org/pdf/2412.06727v1

Refer to caption

Notations	Definition
$D$ $(\cdot)$	Victim detector with output as forgery probability
$N$	The number of particles
$K$	The max number of iterations
$P^{k}$	The $k$ -th iteration of the particle swarm
$w^{k}$	The inertia of the $k$ -th iteration
$[W_{\text{min}},W_{\text{max}}]$	The range of inertia
$C_{\text{p}}$ , $C_{\text{total}}$	Individual confidence and particle swarm confidence
$Pos$ $(\cdot)$	The position of a particle
$Pos_{\text{best}}$ $(\cdot)$	Optimal position of one particle or particles
$f_{i}^{k}$	Fitness of the $i$ -st particle of the $k$ -st iteration
$f_{i,\text{best}}$ , $f_{\text{best}}$	Optimal fitness of one particle / particles
$\alpha,\beta$	The radius and variance of Gaussian blur
$\phi$	The compressive strength of JPEG compression
$\sigma$	The variance of Gaussian noise
$(\lambda_{x},\lambda_{y})$	The coordinates of light spot center
$\theta,\gamma$	The center brightness and radius of light spot