Abstract：With the wide application of deep learning method in speech recognition system，especially in autopilot and personal identification，the security of speech recognition system is crucial.In recent years，the application of deep learning has brought more convenient training steps to speech recognition system，higher recognition accuracy with the potential risks to the security problem.Recent studies have shown that deep neural networks are vulnerable to adversarial attacks in the form of subtle perturbations added onto the input data，resulting in incorrect predictive results.If the speech recognition system based on deep learning is attacked by additional minor disturbances，the autopilot will be attacked by malicious voice attacks，which will bring great security risks to the autopilot system.Aiming at the security of speech recognition system，in this paper we propose a novel blackbox adversarial attack toward speech recognition system，which uses the cuckoo search algorithm to automatically generate the adversarial speech examples to achieve the target attack.Using the generated speech examples to attack the speech recognition system，it is found that there are security vulnerabilities even in the current stateoftheart speech recognition systems.Extensive experiments are carried out on public voice data set，Google voice command data set，GTZAN data set and LibriSpeech data set，to testify the effectiveness of the proposed blackbox attack method.Furthermore，the generated adversarial examples are applied to attack other speech recognition system to testify the strong transfer attack capacity and makes a subjective evaluation test on the them to explore their concealment.
2020, Vol. 41 
