Based on Generative Adversarial Networks Seed Generation Method for Fuzzing

Fuzzing, as one of the effective methods for vulnerability detection, has shown significant impact in discovering potential software vulnerabilities. However, the quality of seed inputs remains a critical factor influencing the effectiveness of fuzzing. and although in recent years, machine-learning based seed generation methods have achieved some results, there are still some problems leading to the failure of test cases generated by fuzzer variants to trigger more security vulnerabilities. This paper employs Base64 encoding and decoding technology to flexibly expand the types of generated seeds. Using the RelGAN model for seed generation and subsequently modifying the generated files with the hot-spot stitching algorithm, we enhance the quality and diversity of the seeds. This significantly improves the performance of fuzzing on target programs with various input formats. The experimental results show that the method in this paper finds a higher number of unique crashes compared to AFL++ using the original seeds.