Deep learning-based solution for smart contract vulnerabilities detection.

Area of Science:

• Blockchain Technology
• Artificial Intelligence
• Software Engineering

Background:

• Smart contracts are vital for decentralized applications but susceptible to vulnerabilities causing financial losses.
• Traditional static analysis tools for smart contract vulnerability detection suffer from high false positives/negatives and lack adaptability.
• Deep learning offers a rule-free approach to learn and identify vulnerability patterns in code.

Purpose of the Study:

• To explore the application of deep learning for detecting smart contract vulnerabilities.
• To introduce Lightning Cat, a deep learning-based solution for vulnerability detection.
• To evaluate the performance of different deep learning models in identifying smart contract flaws.

Main Methods:

• Trained three deep learning models: Optimized-CodeBERT, Optimized-LSTM, and Optimized-CNN.
• Utilized CodeBERT pre-training for accurate code syntax and semantic analysis.
• Extracted vulnerable code function segments to retain critical vulnerability features.
• Evaluated models on the SolidiFI-benchmark dataset comprising 9369 vulnerable contracts.

Main Results:

• The Optimized-CodeBERT model within the Lightning Cat framework achieved the highest performance.
• Optimized-CodeBERT attained an f1-score of 93.53%, surpassing other tested deep learning models.
• The approach demonstrated effectiveness in detecting seven different types of smart contract vulnerabilities.

Conclusions:

• Deep learning, particularly the Optimized-CodeBERT model, presents a superior method for smart contract vulnerability detection compared to static analysis.
• The Lightning Cat solution effectively leverages deep learning for robust and adaptable vulnerability identification in blockchain applications.
• The findings highlight the potential of AI in securing the rapidly evolving landscape of smart contracts.