Research On Context-aware Attacks Of Smart Device Voice Assistants

Voice interaction via voice controlled systems(VCS)is gaining popularity as intelligent devices become more powerful.Voice Controlled systems uses the speech recognition technology to analyze the voice,understand the user's needs,and perform human-computer interaction(HCI)with users in the form of voice assistants.While satisfying users' various application needs in a new interactive way,the system also gains more and more application permissions and can access sensitive system resources.In this process,it may be used by criminals,causing the device to risk information leakage,causing unnecessary losses to users.The thesis focuses on the context-aware attack of the intelligent device voice assistant.Through reading a large number of documents,we understand speech recognition,speech synthesis,context-awareness and related attack methods.On the one hand,it is hoped that the research will inform the public of the existence of such attacks against smartphone voice systems,so that the public can understand the details of the attacks,so as to raise awareness of the prevention of such attacks,and hope to prevent similar attacks from occurring.On the other hand,for this kind of similar attack model,several strategies can be borrowed to defend against this attack model,and contribute to maintaining a harmonious and healthy application environment of intelligent devices.In the research work,in the early stage of this thesis,by investigating the literature of speech recognition,speech synthesis,context-awareness and speech assistant related attacks,the thesis propose a context-aware attack model based on machine learning.This model is mainly composed of speech synthesis module and context-awareness module.It uses speech synthesis model to synthesize speech and combines context-awareness module to achieve attack.For the speech synthesis module,the thesis studies the end-to-end speech synthesis model and the speech concatenation model,and uses the existing model to realize the activation voice.For the context-awareness module,the thesis researches the context-aware technology of machine learning,use the built-in sensors of the smartphone to collect the original data,and then carries out the model after data cleaning and related feature extraction.Aiming at the context-aware attack problem,this thesis uses four machine learning models,decision tree,random forest,XGBoost and logistic regression.The experimental results show that the random forest model is relatively effective in identifying the attack time.Then,based on the real situation experiment,the model also obtained a higher success rate,which basically conforms to the original intention of the design.Finally,this thesis analyzes the limitations of the model by deeply studying the implementation details of the attack model.Through reading a large number of related literatures,three defense strategies,such as double identification authentication method,voice replay attack detection method and man-machine sound source detection method,are sorted out for this similar attack mode.From the perspective of system design,two corresponding defense strategies are proposed,hoping to completely prevent the occurrence of similar attacks and ensure the safety of information and property of the masses.