Hotword Detection Model Audio Attack Prevention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Internet-connected devices capable of receiving and executing voice commands are vulnerable to malicious attacks, where attackers can issue fraudulent commands by outputting recognized hotwords and commands through the device's speaker, leading to potential risks such as data theft or physical danger, as existing audio 'eraser' functionalities do not completely eliminate audio output and thus may still detect fake voice commands.
Innovation Solution
Implementing an eraser hotword model that analyzes audio output data to detect and block fraudulent hotwords, preventing the execution of subsequent voice commands, and communicating with other devices to inform them of fraudulent hotwords, thereby extending security measures across the network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If audio 'eraser' functionality is used to dampen audio output, then the amount of audio received by the microphone is reduced, but the audio output is not completely removed and fraudulent commands can still be detected
Solution Approach 1:
Instead of only analyzing the microphone input to detect hotwords, the patent inverts the approach by analyzing the speaker output to detect fraudulent hotwords. The system monitors what the speaker outputs and identifies when a hotword is being played back, which indicates a potential attack. This inverted detection method complements the traditional microphone-based detection and blocks fraudulent commands even when audio eraser functionality is present.
2Device complexity
If a single hotword model is used for both input and output detection, then device complexity is reduced, but the precision of detecting fraudulent hotwords may be compromised
Solution Approach 1:
The patent applies local quality by creating specialized hotword detection models optimized for specific functions: an input hotword model for detecting legitimate hotwords from microphones and an output hotword model for detecting fraudulent hotwords from speakers. Each model is trained and configured for its specific detection context, ensuring high precision. This specialization allows the system to maintain high detection accuracy while clearly separating the detection functions.
Data Source
AI summary
In some implementations, a method includes receiving output audio data that is provided to a speaker of a device and that represents audio for output by the device, receiving, after the output audio data is provided to the speaker of the device, input audio data that represents audio detected by a microphone of the device, determining, by an output hotword detection model, that the output audio data that is provided to the speaker of the device includes a representation of a hotword, determining, by an input hotword detection model that is less accepting of hotwords than the output hotword detection model, that the input audio data that represents audio detected by a microphone of the device includes a representation of a hotword, and, in response, blocking use of the input audio data to initiate a command.