HSS-GAT: A Memory-Efficient Hybrid Structural and Semantic Learning Using Graph Attention Networks and TF-IDF Features for AI-Generated Text Classification

Abstract

The rapid growth of content generated by large language model (LLM)-based applications has posed significant challenges to the society especially in domain where content originality, authorship and factual accuracy are critical. Therefore, there is an increasing need to develop robust and reliable classification methods for classifying AI-generated text. Such methods, can help this will help maintain information credibility, preventing spreading misinformation, detecting academic plagiarism, protecting digital platforms from malicious content, and ensuring transparency in online communication. Although, researchers have introduced plenty of detection methods and frameworks. there is still a need of more reliable solutions due to rapid advancement of LLM models which are capable of generating content that closely resemble human-written contents. The proposed framework presents the hybrid graph-based AI-generated text classification approach (HSS-GAT) that combines structural graph learning with TF-IDF based global semantic features to improve the classification performance. The experiment was conducted using two benchmark datasets (HC3 and DAIGT v2) which includes the generated contents from diverse LLM models. The experimental result on combined (HC3 + DAIGT v2) dataset demonstrates that the proposed framework achieved an accuracy of 98.15% and an F1-Score of 97.93% highlighting the effectiveness of the proposed framework.