Neuron Activation Frequency Based Test Case Prioritization

Abstract

Deep neural networks (DNNs) have been increasingly adopted in various applications. Systematic verification and validation is essential to guarantee the quality of such systems. Due to the scalability problem, formal methods can hardly be widely applied in practice. Testing is one of feasible solutions. However, lacking of input space specification for DNNs requires a large set of test cases to be constructed to increase the testing adequacy, which leads to high labeling cost of test cases. In this paper, we put forwards a test case prioritization method for DNN classifiers, which assigns high priorities to those cases that could lead to wrong classifications. The priorities are calculated according to the activation pattern of neurons acquired from the training sets and the activated neurons collected from certain inputs. For a trained model, the method consists of two steps. First, we accumulate neuron activation patterns over the training set, and construct a set of frequently activated neurons based on the frequency (times) of activation for every class. Second, the metrics are computed according to the comparison between the activated neurons of an input and the selected set of frequently activated neurons with its output. The experimentation is carried out over three popular datasets with various neural network structures. The results demonstrate that the test cases with higher priorities are more prone to be mis-classified. And the prioritized test cases over a DNN model within same datasets are also efficient in triggering mis-classification of other DNNs with similar structures.