RGChaser: A RL-guided Fuzz and Mutation Testing Framework for Deep Learning Systems

Due to the powerful performance of deep learning (DL) systems, they have been increasingly deployed in numerous safety-critical domains. However, DL systems are known to be vulnerable to adversarial examples. Although testing techniques have been applied to ensure the robustness of DL systems nowadays and made gratifying progress, existing fuzz testing for DL systems does not pay attention to more diverse predetermined goals, nor does it provide guidance on how to mutate the seeds. Moreover, existing mutation testing for DL systems also suffers from similar troubles. To fill this gap, we investigate fuzz testing and mutation testing guided by reinforcement learning (RL) for DL systems. Meanwhile, we propose RGChaser, a RL-Guided testing framework which combines these two techniques to generate and screen more diverse high-quality test data for chasing potential vulnerabilities. Besides, to further improve the diversity of the generated mutants, we design a set of mutation operators for convolutional neural network (CNN) and continuous mutation spaces. We have implemented the introduced techniques on the MNIST and CIFAR-10 datasets to evaluate their practicality. Experimental evidences show that: (1) our approach is useful for generating diverse test suites efficiently, by upto 71.6% reduction in required mutations and 10 more success rates; (2) our approach could provide specific guidance on how to generate goal-based mutants.