Using Transformer and GAN Models for Software and Security Testing

Abstract

This paper investigates how Generative Adversarial Networks (GANs) and transformer models can support the process of software and security testing by generating and augmenting test data. We start with an analysis of the use of GANs for software testing, focusing on the generation of privacy-sensitive data in the automotive domain. We demonstrate that GANs can contribute to the efficient generation of test data which meet specific technical and regulatory requirements and discuss the limitations of applying differential privacy in this context. Based on this intermediate result, we investigate how lightweight open source transformer models can be applied to fuzzing to detect weaknesses. The evaluation is carried out using a modular training and evaluation framework. Our system implements the “Beyond Random Inputs” fuzzing approach by Rostami et al. (2024), using the Lua interpreter1as the fuzzing target. We then compare its effectiveness with the coverage-guided fuzzer AFL++ in terms of code coverage and vulnerability detection. Our results demonstrate the potential and limitations of transformer-based fuzzing in constrained environments, motivating further research on model scaling, resource efficiency, and domain transferability.