DLRegion: Coverage-guided fuzz testing of deep neural networks with region-based neuron selection strategies

Abstract

Deep Learning (DL) systems have been increasingly applied to safety-critical scenarios, such as autonomous driving or medical diagnoses. However, it often exhibits erroneous behaviors that can cause serious consequences. Therefore, it is important to ensure the quality of the DL systems through systematically testing. Recent research works have proposed many testing techniques for deep neural networks (DNNs), among which the coverage-guided fuzz testing has achieved remarkable results. The neuron selection strategy is the key ingredient of the technique. It can affect the technique’s effectiveness. However, current neuron selection strategies did not utilize the output distribution of each neuron on the training data, which can characterize the behavior of the neuron. This paper introduces DLRegion, a coverage-guided fuzz testing technique of DNNs with region-based neuron selection strategies. DLRegion can expose erroneous behaviors of DNNs while maximizing coverage. DLRegion first incorporates a seed selection strategy based on the level of confidence in the classification of the inputs to select seed inputs to mutate. DLRegion also proposes region-based neuron selection strategies, which utilize the region where the output value of each neuron is in its output distribution to select valuable neurons to activate for covering more internal states. Empirical studies on three well-known datasets, guided by five existing criteria demonstrate that: (1) the proposed seed selection strategy very effectively improves coverage and defect detection; (2) selecting neurons in different regions has obvious differences in achieving model coverage and detecting defects; (3) DLRegion outperforms other existing techniques by coverage under different criteria and effectively identifies the quantity and diversity of defects; (4) the effectiveness of DLRegion in improving the model robustness. DLRegion not only can cover more internal logic of the DNNs but also can effectively detect DNNs’ erroneous behaviors. Moreover, DLRegion can improve the robustness of the model.