ViScheduler: Visualized Seed Scheduling Based on Runtime Features
Hang Xu, Jianshan Peng · 2024
Fuzzing is a popular technique for software vulnerability detection. Modern fuzzing uses feedback from the program under test (PUT) to prioritize test cases, but existing seed scheduling methods do not provide an interface for expert intervention. We propose a method to visualize seed features using neural networks trained with runtime seed features. This allows human experts to visualize fuzzing progress and intervene in seed scheduling. Current challenges include building labeled machine learning datasets using runtime features and visualizing high-dimensional seed embeddings. We address these issues by proposing a Transformer neural network trained using an approximation task (bug test case embedding extraction). The feature embeddings are then visualized in 3D space using dimensionality reduction methods. We also introduce four seed scheduling strategies based on seed embeddings to facilitate human intervention. This approach bridges the gap between seed scheduling and human expertise and experimental results show that it can lead to more effective fuzzing.