Enhancing Security and Performance in PyTorch: A Hybrid Fuzzing Approach

As software systems become more intricate, the prevalence of memory safety vulnerabilities poses a growing concern for security. This study delves into the realm of soft-ware security, focusing on PyTorch, a popular machine-learning framework. Leveraging hybrid fuzzing techniques, the research aims to identify and address memory safety issues in PyTorch, emphasizing its significance in the context of AI applications. The study introduces improvements to the sydr-fuzz tool, optimizing its security predicates verification and implementing a novel scheduling strategy for memory modeling. The findings include the discovery of multiple security vulnerabilities in PyTorch and demonstrate notable enhancements in the performance of the hybrid fuzzing tool.