notesum.ai

Published at November 18

Exploring adversarial robustness of JPEG AI: methodology, comparison and new methods

eess.IV

cs.AI

cs.CV

Released Date: November 18, 2024

Authors: Egor Kovalev¹, Georgii Bychkov¹, Khaled Abud¹, Aleksandr Gushchin², Anna Chistyakova³, Sergey Lavrushkin², Dmitriy Vatolin², Anastasia Antsiferova⁴

Aff.: ¹Lomonosov Moscow State University; ²MSU Institute for Artificial Intelligence; ³ISP RAS Research Center for Trusted Artificial Intelligence; ⁴Laboratory of Innovative Technologies for Processing Video Content, Innopolis University

Arxiv: http://arxiv.org/abs/2411.11795v1

Optimisation target	Formula
FTDA default	$\left\\|C(x)-C(x^{\prime})\right\\|_{2}$
Added-noises	$\left\\|C(x^{\prime})-C(x)-(x^{\prime}-x)\right\\|_{2}$
Reconstruction	$\left\\|C(x^{\prime})-x^{\prime}\right\\|_{2}$
FTDA MS-SSIM	MS-SSIM $\left(C(x),C(x^{\prime})\right)$
Reconstruction MS-SSIM	MS-SSIM $\left(x^{\prime},C(x^{\prime})\right)$
BPP increase	$1-bpp\left(C(x^{\prime})\right)$
Y-modification of all targets	Take only Y in YCbCr