Core Concepts
Leveraging style transfer to create natural and undetectable adversarial images.
Abstract
Adversarial attacks in Virtual Reality pose security threats.
Proposed framework uses style transfer for natural adversarial images.
Utilizes a latent text-to-image diffusion model for image generation.
Incorporates neural style transfer and an adversarial attack network.
Evaluation includes qualitative and quantitative assessments.
Stats
T-shirt, 93.67
Umbrella, 96.05
Sleeping bag, 86.54
Zebra, 93.59
NIMA ↑ (0∼10)
Topiq iaa ↑ (0∼10)
Topiq nr ↑ (0∼1)
Tres ↑ (0∼100)
Quotes
"Our approach successfully generates naturalistic adversarial images while maintaining competitive attacking performance."
"We provide a novel non-reference perceptual image quality assessment method."
"Our Diffusion Attack is able to achieve higher image quality and aesthetic assessment average scores compared with baselines."