Deep learning (DL) systems are being used in scenarios with increasingly complex environments, especially in human security scenarios, such as autopilot, medical diagnosis and facial recognition. These applications with DL systems inside not only bring great convenience, but also bring severe challenges. The fault behavior of DL systems will not only lead to huge economic losses, but even human casualties. Therefore, making software testing to ensure the quality of such systems particularly important. Traditional software testing techniques are not suited to the internal structure and operating mechanism of DL systems. Traditional fuzz testing commonly exist inefficiencies such as invalid test case selection and generation leading to waste of resources. Recent fuzz testing methods did not accurately utilize the output of key neurons of critical neuron layers, which could not be more efficiently affect adversarial input in the terms of test case screening accuracy. This paper proposes an adaptive key neuron-guided method with fuzz testing for test case selection. The algorithm developed based on the internal feature of deep neural network. Empirical investigations have been undertaken to assess the efficacy of the suggested methodology utilizing two distinct datasets, one fuzzer, four test case selection algorithms and five DL models. The results show that the strategy can significantly improve the efficiency of the state-of-the-art fuzzer for DL models, not only in terms of screening more adversarial inputs compared with other test case selection algorithms, but also achieving higher mutation and generation success rate.