Method for defending inaudible instruction to control voice assistant based on machine learning

Abstract

The invention discloses a method for defending an inaudible instruction to control a voice assistant based on machine learning. The method comprises the following steps: 1) acquiring the data of positive and negative samples, 2) performing data segmentation on the positive and negative sample data; 3) performing sample feature selection and normalization; 4) selecting a classifier for training andgenerating a detection model used for detecting a malicious voice command; and 5) detecting the voice command to be detected through the obtained detection model. According to the invention, a creative feature selection method is selected, and the method is directed to intelligent devices with different models, and a normal voice command and the malicious voice command need to be acquired throughthe intelligent devices of the type, and a specific classifier for the type of device is trained as positive and negative samples. Through the customized mode, the problem that the detection and defense cannot be carried out by cross devices can be well solved.

Description

technical field [0001] The invention belongs to the security field of artificial intelligence voice assistants, and in particular relates to a method for controlling voice assistants based on machine learning defense silent commands. Background technique [0002] As a convenient and effective method of human-computer interaction, voice assistants such as Siri, Google Now, Alexa, Cortana, S Voice, and Hi Voice are becoming more and more popular in people's lives. Almost all smart devices (smart phones, tablets Computers, wearable devices and smart speakers) are equipped with corresponding voice assistants. [0003] However, voice assistants also face various security risks, such as a method called "dolphin sound attack" [Zhang, G., Yan, C., Ji, X., Zhang, T., Zhang, T.,&Xu,W.(2017).DolphinAtack:Inaudible Voice Commands.arXiv preprint arXiv:1708.09537.], using the vulnerability of the microphone circuit of the smart device, the voice assistant can be silently controlled to or...

AI Technical Summary

Problems solved by technology

[0009] 5. Hidden attack: The screen display and voice feedback may expose the attack, but the hacker can reduce the screen brightness or volume to achieve the purpose of hiding

[0018] 1. Manufacturers are unable to improve the hardware of smart devices that have already been sold, or the cost of recall is unaffordable for manufacturers

[0019] 2. The ability of the microphone circuit to receive ultrasonic waves has always been a problem. Although with the progress and development of technology and technology, it has not been effectively solved. Therefore, the defense scheme based on hardware is a high-cost, difficult and long-term defense. solutions, cannot effectively solve the existing problems

[0020] 3. The software-based solution in the prior art only tested a voice command "Hey" on one smart device, so it is not sure whether this method can work on different smart devices, different voice commands, and different human voices.

Images