A FEW-SHOT COPYRIGHT ATTRIBUTION FOR GAME IMAGE BASED ON META-LEARNING

Abstract

The rapid proliferation of digital game content has increased the need for reliable methods to identify the copyright ownership of game images, particularly in scenarios where only a limited number of labeled samples are available for each title, character, or visual asset. Conventional image classification approaches often require large-scale annotated datasets and show limited generalization across heterogeneous game domains, such as variations in genre, art style, rendering quality, and post-processing. To address these challenges, in this paper, we propose a few-shot copyright attribution framework for game images based on meta-learning. The proposed method is designed to learn transferable visual representations and fast adaptation strategies from episodic training tasks, enabling effective attribution of previously unseen or weakly represented game image classes. In addition, a cross-domain learning setting is considered to improve robustness against stylistic and content-level discrepancies among game datasets. The framework aims to distinguish copyright-related visual sources by capturing discriminative fine-grained patterns while maintaining generalization under data scarcity. In this paper, we have simulated a few-shot attribution scenario on a game image benchmark and evaluated the proposed framework using standard recognition and attribution measures. The results are expected to demonstrate that meta-learning provides a practical and scalable solution for copyright attribution in game images, offering strong potential for digital copyright protection, content monitoring, and forensic analysis in the gaming industry.