Demystifying the Vetting Process of Voice-controlled Skills on Markets
summary
Paper reverse-engineers skill vetting and shows concrete weaknesses: a relevance-based Turing test using ConceptNet distinguishes human vs machine vetters; 1.9% of Amazon and 55.8% of Google vetter IDs appear human, machines start vetting ~222 seconds after submission, Amazon runs eight userIDs in sequence while Google uses one. Over an 8-month run with >200 submitted skills, hidden voice commands never get exercised by machines. Three concrete attacks and a set of 30 attack sentences reliably confuse NLP-based parsing and bypass automated checks.
AI-written summary · about
related talks
26mSkill Squatting Attacks on Amazon Alexa
39m Put Words In My Mouth
33mYour voice confirms my identity
48mSecurity Vulnerabilities in AI Assistant Based Applications
05mExploring Authentication for Security-Sensitive Tasks on Smart Home Voice Assistants
42mIs your AI plotting to kill you? Why AI and Machine Learning is dangerous
30mHidden Voice Commands
19mTesting LLM-Powered Applications