DeepHC: Efficient Generating Tests with High Coverage for Deep Neural Networks.

As deep neural networks (DNNs) are increasingly used in the software of aerospace, aviation, and other safety-critical fields, ensuring their reliability and safety becomes paramount. Software testing is still the most critical means to ensure software quality. In order to guarantee the adequacy and effectiveness of testing, aerospace and other industrial areas usually require very high test case coverage. Despite considerable advances in structured coverage criteria and coverage-guided adversarial examples generation for DNNs, there is a lack of in-depth research on how to efficiently generate test inputs with high coverage, which is essential for software testing. In this paper, we propose DeepHC, a general framework combining fuzzing and gradient-based algorithms to achieve high coverage towards 6 most widely used coverage criteria. To explore test inputs with high coverage, DeepHC adopts a novel evaluation strategy for selecting inputs that can activate neurons more efficiently. We evaluate the effectiveness of DeepHC on 2 popular datasets and 3 DNN models with different complexities. The experimental results show that DeepHC outperforms the state-of-the-art method in enhancing coverage.