The present disclosure discloses a machine learning-based method for defending a voice assistant from being controlled by an inaudible command, including following steps: 1) collecting data of positive and negative samples, 2) performing data segmentation on data of the positive and negative samples; 3) selecting and normalizing sample features; 4) selecting a classifier to be trained and generate a detection model for a malicious voice command; 5) detecting a voice command to be detected by the detection model. The present disclosure selects an original feature selection method, and for smart devices of different types, it is necessary to obtain normal voice commands and malicious voice commands by means of a smart device of this type, and use them as the positive and negative samples to train a specific classifier for the device. Such a customized approach can well solve a problem that detection and defense between devices cannot work.