DRFuzz: A Regression Fuzzing Framework for Deep Learning Systems
DRFuzz is a novel regression fuzzing framework for deep learning systems. It is designed to generate high-fidelity test inputs that trigger diverse regression faults effectively. To improve the fault-triggering capability of test inputs, we adopt a Markov Chain Monte Carlo (MCMC) strategy to select mutation rules that are prone to trigger regression faults. Furthermore, to enhance the diversity of generated test inputs, we propose a diversity criterion to guide triggering more faulty behaviors. In addition, DRFuzz incorporates a GAN-based fidelity assurance method to guarantee the fidelity of test inputs. We conducted an empirical study to evaluate the effectiveness of DRFuzz on four regression scenarios (i.e., supplementary training, adversarial training, model fixing, and model pruning). The experimental results demonstrate the effectiveness of DRFuzz.