Plagiarism Detection in GAN-Generated Abstract Art: A Multi-Modal Semantic and Compositional Approach

This paper presents an interpretable multi-modal framework for screening potential plagiarism in GAN-generated abstract art. Because abstract works often resemble one another through palette, texture, rhythm, and massing rather than recognizable objects, single-metric or text-oriented plagiarism tools are insufficient. The proposed pipeline combines perceptual cues (MS-SSIM, color-distribution distances, Gram-matrix texture statistics, and edge topology), compositional cues (symmetry, balance, saliency spread, orientation entropy, and palette harmony), and semantic cues from CLIP and BLIP. Each channel is normalized, fused into a calibrated similarity score, and reported with uncertainty bounds and channel-level explanations. Using representative WikiArt-anchored cases and GAN-generated counterparts, the framework distinguishes probable derivation, stylistic influence, and independent creation more reliably than any isolated metric. The revised manuscript adds a consolidated related-work matrix, documented case provenance for A1–A5, illustrative output dossiers, and visual summaries of the comparative results. The method is intended as a transparent decision-support tool for scholarly, curatorial, and legal review rather than an automated adjudicator.