Research on Deep Learning Network Sample Generation Method Based on Abnormal Examples

The complexity and uncertainty of deep neural networks (DNN) necessitate comprehensive behavior testing for model quality assurance. Current studies generating samples via coverage criteria and fuzz testing often neglect sample diversity and error-detection ability. The surprise adequacy metric, quantifying neuron output differences between samples and the training set, lacks a specific generation method. Herein, a novel method is introduced. Initially, key decision-influencing neurons are identified to refine the surprise adequacy metric. Subsequently, seed samples with error-revealing potential are filtered by this metric. Finally, integrating coverage-guided fuzz testing, with the surprise adequacy value and class probability difference as joint optimization targets, the gradient ascent algorithm computes perturbations for iterative sample generation. Five DNN models and four image datasets were used for validation. Results confirm that the enhanced metric swiftly captures abnormal samples and saves computation time. In sample generation, compared to DeepGini and RobOT, the derived test sets' surprise coverage rises by up to 5.9% and 15.9% respectively; versus DLFuzz and DeepXplore, the proposed method achieves a boost of up to 26.5% and 33.7%, validating its efficacy.