비가시성 워터마킹과 선택적 클래스 언러닝을 통합한 이미지 생성 프레임워크 = An Image Generation Framework Integrating Invisible Watermarking and Selective Class Unlearning

다국어 초록 (Multilingual Abstract)

Recent rapid advancements in generative models have enabled the automatic creation of high-quality content but have also raised significant social concerns, such as copyright infringement, privacy leakage, and the generation of harmful content. Consequently, the necessity for Machine Unlearning, which removes the influence of specific data post-training, and Invisible Watermarking, which verifies the provenance and authenticity of generated content, has become critical. However, existing research has treated these technologies independently. Therefore, in this paper, we propose a novel image generation framework that integrates invisible watermarking and selective class unlearning. The proposed method ensures invisibility and robustness using a hybrid watermarking technique combining Discrete Cosine Transform and Singular Value Decomposition, and effectively forgets target classes through a two-step learning strategy that induces confusion in the discriminator. Experimental results demonstrate that the proposed method achieves effective suppression, while maintaining original-level quality for remaining classes. Furthermore, the stable detection of watermarks after unlearning confirms that the proposed framework contributes to building safe and trustworthy generative models.