<![CDATA[The rapid advancement of generative artificial intelligence (AI) has made synthetic digital paintings increasingly difficult to distinguish from human-made artworks, raising concerns regarding authenticity, copyright protection, and digital forensics. The main objective of this research is to develop a reliable and interpretable framework for distinguishing AI-generated paintings from human-created artworks by integrating visual and noise-based features. To address the limitations of conventional single-stream CNN models, this study proposes a Two-Stream Network with a Late Fusion strategy, combining a visual stream based on EfficientNetV2-S and a noise stream based on Xception with Spatial Rich Models (SRM).The proposed architecture processes semantic visual features and residual noise characteristics independently, followed by weighted decision-level fusion with a ratio of 0.7:0.3. Experiments were conducted using the AI-Artwork public dataset from Kaggle, consisting of 15,000 images with a data split of 64% training, 16% validation, and 20% testing. Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC, ensuring a comprehensive assessment beyond accuracy alone. The results demonstrate that the proposed method achieves 98% accuracy, 98% precision, a 99% F1-score, and high discriminative capability compared to single-stream baselines. Model interpretability was analyzed using Grad-CAM to examine the contribution of each stream. Despite promising results, this study is limited by evaluation on a single dataset and static fusion weights, which may affect generalization to unseen generative models. Future work includes cross-dataset evaluation, adaptive fusion strategies, and exploration of lightweight architectures. Practically, this approach has potential applications in digital art authentication, forensic analysis, and content moderation systems, as well as supporting emerging policies for AI-generated content regulation and copyright protection.]]>
