DeepLogic: Priority Testing of Deep Learning Through Interpretable Logic Units

Abstract

With the increasing deployment of deep learning-based systems in various scenes, it is becoming important to conduct sufficient testing and evaluation of deep learning models to improve their interpretability and robustness. Recent studies have proposed different criteria and strategies for deep neural network (DNN) testing. However, they rarely conduct effective testing on the robustness of DNN models and lack interpretability. This paper proposes a new priority testing criterion, called DeepLogic, to analyze the robustness of the DNN models from the perspective of model interpretability. We first define the neural units in DNN with the highest average activation probability as “interpretable logic units”. We analyze the changes in these units to evaluate the model's robustness by conducting adversarial attacks. After that, the interpretable logic units of the inputs are taken as context attributes, and the probability distribution of the softmax layer in the model is taken as internal attributes to establish a comprehensive test prioritization framework. The weight fusion of context and internal factors is carried out, and the test cases are sorted according to this priority. The experimental results on four popular DNN models using eight testing metrics show that our DeepLogic significantly outperforms existing state-of-the-art methods.