In recent years, various attack transactions have become one of the most destructive security threats in the Ethereum ecosystem, causing significant financial losses. The complex dynamics of attack transactions often require learning-based detection methods to employ structurally complex models, resulting in substantial modeling overhead. Meanwhile, scarce attack samples provide limited supervision, constraining both detection efficiency and effectiveness. To address these problems, this paper proposes an Efficient Lightweight graph modeling with large language model (LLM)-enhanced Supervision for Ethereum Attack transaction detection (ELSA). First, to improve the efficiency of modeling transaction dynamics, ELSA designs a transaction graph structure based on function call traces. This structure enables lightweight modeling of transaction behavior and improves model efficiency. Second, to address the lack of supervision caused by scarce attack labels, ELSA constructs supervisory signals from the comment of the first function triggered by a transaction. ELSA further introduces the large language model Qwen2.5-Coder:7B to refine the comment. The refined comment provides high-quality supervision and improves detection performance. Experimental results show that, compared with state-of-the-art methods, ELSA reduces training time, trace storage footprint, and per-block data collection time to approximately 15%, 4%, and 13% of the SOTA level, respectively. Under fully supervised and few-shot settings, the proposed method improves the overall AUC by 0.03 and 0.02 absolute points, respectively, compared with state-of-the-art methods.
