GPT-Based Wasm Instruction Analysis for Program Language Processing

GPT-Based Wasm Instruction Analysis for Program Language Processing

연구 분야: Artificial Intelligence

학회: International Symposium on Emerging Information Security and Applications

Capturing the logical structure of programming languages poses a significant challenge for program analysis. Given the complex syntax rules, subjective code vulnerabilities, irrelevant statements, code annotations, and intricate structural information, related studies have explored various semantic comprehension and intermediate representation approaches to extract precise information for program analysis. However, most research in the generic domain ignores defective and non-defective program code, putting them in the same category. In this paper, we introduce a new program analysis method that combines WebAssembly (Wasm) instructions with a 20-billion-parameter transformer model and natural language processing. This approach aims to advance the capabilities of program analysis tools in computer science by jointly embedding Wasm instructions and natural language for more effective program analysis. Our experiments demonstrate that this fused embedding approach achieves state-of-the-art performance, and the accuracy reaches approximately 98 percent, better than traditional small-scale weight models based on intricate conversion tasks such as abstract syntax trees(AST). Moreover, it is more valuable to classify potential vulnerable and non-vulnerable programs in the formal verification special field. Our exploration enhances traditional program classification methods in software security and introduces the application of GPT to offer a more straightforward, convenient, and high-performance approach.