White-Box Test Input Generation for Enhancing Deep Neural Network Models through Suspicious Neuron Awareness

Deep Neural Network (DNN) testing has emerged as an effective way of uncovering erroneous behaviors in DNN models and further enhancing their performance. Research on test input generation has gained much attention from both researchers and practitioners, aiming to expose faults in models. The newly generated inputs subsequently serve as additional training instances for model refinement through retraining. Existing approaches generate test inputs by optimizing an objective function based on testing metrics such as neuron coverage and property-related metrics, and the gradient of the objective is used to perturb seed inputs. However, these approaches pay limited attention to the model’s decision logic, particularly the erroneous decision patterns learned during training. Furthermore, they primarily focus on detecting faults without considering the diversity of detected misbehaviors, which limits the models’ ability to learn diverse features through retraining. To address these limitations, this article introduces SUNTest, a novel test input generation approach designed to detect diverse faults and enhance the robustness of DNN models. SUNTest focuses on erroneous decision-making by localizing suspicious neurons responsible for misbehaviors through the execution spectrum analysis of neurons. To guide input mutations toward inducing diverse faults, SUNTest designs a hybrid fitness function that incorporates two types of feedback derived from neuron behaviors, including the fault-revealing capability of test inputs guided by suspicious neurons and the diversity of test inputs. Additionally, SUNTest adopts an adaptive selection strategy for mutation operators to prioritize operators likely to induce new fault types and improve the fitness value in each iteration. Experiments conducted on eight DNN models demonstrate the effectiveness of SUNTest in fault localization and test input generation. It outperforms existing test input generators in the number of detected faults, uncovering up to 80.9 more distinct fault types. In terms of model enhancement, SUNTest increases the average accuracy improvement by up to 8.04% compared to baseline approaches.