A Method of Generating Test Data for Intelligent System Based on Uncertainty

Abstract

A method for generating test data of an intelligent system based on uncertainty provided by the present invention reconstructs the traditional CNN classifier model, combines the uncertainty reasoning of D-S evidence theory to obtain the information conflict value, and then based on the conflict value and Neuron coverage guides test data generation. From the perspective of generating more diverse test data and improving test adequacy, by comprehensively increasing the conflict value between sample features, adversarial samples can be generated to induce model misclassification and improve neuron coverage, which can help improve test adequacy The content of the two aspects, combining the structure of the model, the parameters, the conflict between the sample features and the neuron coverage index, is a new test data generation method. Compared with DeepXplore, there is no need to use multiple neural network models for cross-reference, and compared with DLFuzz, more test data can be generated.

Description

technical field [0001] The invention relates to the technical field of media communication, in particular to an uncertainty-based intelligent system test data generation method. Background technique [0002] The adversarial-based test data generation method essentially uses adversarial sample generation techniques to generate adversarial samples that cause the model to misclassify, thereby exposing the model to defects. The following first summarizes the adversarial sample generation technology: [0003] Adversarial samples refer to test samples that cause the model to erroneously output with high confidence by deliberately adding some subtle disturbances that cannot be detected by humans to the original data set. In recent years, a variety of adversarial example generation algorithms have been produced according to different optimization methods. According to the perturbation of full pixel and partial pixel perturbation, from the research of perturbation of full pixel, Sz...

AI Technical Summary

Problems solved by technology

[0011] None of the aforementioned adversarial-based test data generation techniques and coverage criterion-based test data generation techniques involve combining the conflicting values ​​obtained in the prediction uncertainty to give a method for generating test data

Images