Structural Test Coverage Criteria for Deep Neural Networks

Abstract

Deep neural networks (DNNs) have a wide range of applications, and software employing them must be thoroughly tested, especially in safety-critical domains. However, traditional software test coverage metrics cannot be applied directly to DNNs. In this paper, inspired by the MC/DC coverage criterion, we propose a family of four novel test coverage criteria that are tailored to structural features of DNNs and their semantics. We validate the criteria by demonstrating that test inputs that are generated with guidance by our proposed coverage criteria are able to capture undesired behaviours in a DNN. Test cases are generated using a symbolic approach and a gradient-based heuristic search. By comparing them with existing methods, we show that our criteria achieve a balance between their ability to find bugs (proxied using adversarial examples and correlation with functional coverage) and the computational cost of test input generation. Our experiments are conducted on state-of-the-art DNNs obtained using popular open source datasets, including MNIST, CIFAR-10 and ImageNet.

Coverage-Guided Testing for Recurrent Neural Networks

“…We comment that with large enough amount of test cases, coverage-guide testing approach provides a new way for the measure and selection of more robust classifier. This is compatible with the results in [39] that a poorly trained neural network exposes more adversarial samples subject to well-defined coverage guided testing. Answer to RQ3: By exploiting the model's internal behaviours, TESTRNN is able to capture the LSTM adversarial samples.…”

“…(Right) Relation between Coverage Metrics. NC: neuron coverage [32], BS: basic state coverage [12], BT: basic transition coverage [12], MC/DC: modified condition/decision coverage [39]. Arrows represent the "weaker than" relation between metrics.…”

“…Although NC and BC have similar formal expressions, BC concerns the boundary value rather than a value that indicates the activation status. For the MC/DC metrics [39], they can be adapted to work between time steps in the new context of RNNs. With such adaptation, MC/DC encourages the exploration of relations between time steps, and therefore are weaker than TC.…”

“…Coverage-guided testing has achieved a great success in software defect detection, and has been extended to work with FNNs in e.g., [32], [48], [25], [39], [40], where a collection of FNN coverage metrics can be found. The definitions of these metrics are based on the structural information, such as the neurons' activations [32], [25], the relation between neurons in neighboring layers [39], [40], etc.…”

“…Coverage-guided testing has achieved a great success in software defect detection, and has been extended to work with FNNs in e.g., [32], [48], [25], [39], [40], where a collection of FNN coverage metrics can be found. The definitions of these metrics are based on the structural information, such as the neurons' activations [32], [25], the relation between neurons in neighboring layers [39], [40], etc. While existing coverage metrics for FNNs may be adapted to work with RNNs, they are insufficient because they do not work with the internal structures of RNNs and, more importantly, the most essential ingredient of RNNs -the temporal relation -is not considered.…”