본문 바로가기
  • Home

Region-Specific Noise Generation for Untransferable Examples Against Neural Style Transfer

  • Journal of Software Assessment and Valuation
  • Abbr : JSAV
  • 2025, 21(4), pp.63~77
  • Publisher : Korea Software Assessment and Valuation Society
  • Research Area : Engineering > Computer Science
  • Received : October 20, 2025
  • Accepted : December 20, 2025
  • Published : December 26, 2025

Park Sunghwan 1 KimJongSeong 1 Byunghoon Oh 2 HWANG YO HAN 3 Junho Hong 4 Jaewoo Lee 1

1중앙대학교
2중앙대
3중앙대학교 산업보안연구소
4성신여대

Accredited

ABSTRACT

Neural Style Transfer poses a significant threat to the intellectual property rights of digital artists, as it can extract and replicate unique artistic styles without consent or compensation. Existing protection techniques face a fundamental trade-off: they either degrade image quality for effective protection or offer minimal protection to preserve quality. This paper proposes a novel region-specific noise generation that solves this trade-off by spatially segregating protection objectives. Our method divides an image into two distinct regions and applies different perturbation strategies to each: a small patch region (2% of image area) receives strong perturbations to hijack the attention mechanism, while the background region (98%) is subjected to minimal perturbations to corrupt global feature statistics. The key innovation lies in the spatial separation that fundamentally eliminates gradient interference between different loss objectives. Experimental results demonstrate that our method achieves over 90% reduction in style transfer effectiveness (STDR 0.8995) while maintaining high visual quality (SSIM 0.8624), representing a 38% improvement in protection efficiency compared to existing methods.

Citation status

* References for papers published after 2024 are currently being built.