Small language models for smart contract vulnerability detection : exploring efficient fine-tuning approaches for reentrancy detection in resource-constrained environments

Abstract

This dissertation investigates the application of parameter-efficient fine-tuning techniques to small language models (1-3B parameters) for detecting reentrancy vulnerabilities in Solidity smart contracts. Despite the prevalence and persistent threat of reentrancy vulnerabilities in blockchain ecosystems, existing detection approaches often require substantial computational resources or extensive manual engineering. This research demonstrates that even modest-sized models can develop meaningful security analysis capabilities when efficiently adapted through Low-Rank Adaptation (LoRA) and trained on synthetic vulnerability data. Experimental results show significant improvements in detection performance after fine-tuning, with accuracy increasing from 48% to 67% for LLaMA 3B and from 45% to 59% for Qwen2.5Coder 3B. This work establishes a foundation for resource-efficient approaches to smart contract vulnerability detection, addressing the gap between the capabilities of frontier language models and practical deployment constraints in development environments. Beyond the specific task of reentrancy detection, the methodology and findings contribute to the broader understanding of efficient language model adaptation for specialized security applications.