LEAST: Loss Ensemble Style Transfer against Arbitrary Style Transfer

  • Park, Sunghwan; 
  • Kim, Jongseong; 
  • Oh, Byunghoon; 
  • Hwang, Yohan; 
  • Lee, Jaewoo; 
  • 외 1명
Citations

WEB OF SCIENCE

0

초록

Arbitrary neural style transfer (AST) models pose a significant copyright threat to artists. While various proactive methods of protection exist, they are somewhat limited; some focus on model ensembles, while others prioritize imperceptibility by constraining perturbations to the high-frequency domain. This high-frequency constraint, however, fundamentally limits attack effectiveness by ignoring low-frequency style components like feature mean. This paper proposes loss ensemble adversarial style transfer protection (LEAST), a novel method that comprehensively attacks several foundational mathematical definitions of style within the VGG feature space (including statistics, Gram matrices, and feature maps) via a composite loss function. Optimized with a full-spectrum gradient, LEAST comprehensively disrupts the VGG feature space, which serves as the de facto standard for style representation. Experiments show that LEAST achieves exceptional JPEG compression robustness (97.5% retention rate, highest among all methods) while demonstrating strong protection against CNN-based AdaIN models and maintaining a balanced performance across diverse architectures.

키워드

Arbitrary Style Transfer; Loss Ensemble; Adversarial Attack; Copyright Content Protection
제목
LEAST: Loss Ensemble Style Transfer against Arbitrary Style Transfer
저자
Park, Sunghwan; Kim, Jongseong; Oh, Byunghoon; Hwang, Yohan; Lee, Jaewoo; Hong, Junho
DOI
10.22967/HCIS.2026.16.059
발행일
2026-10
유형
Article
저널명
Human-centric Computing and Information Sciences
권
16