Security Vulnerabilities in AI Assistant Based Applications
summary
Claims that malware can record assistant commands and location and an app was removed from the Play Store; a physical attack using an inductive speaker demo fails at the conference but the technique can turn windows into loud speakers; voice splicing lets attackers reconstruct “OK Google” and OK Google splicing works intermittently while 30 second voice cloning is plausible; business logic flaws let attackers add contacts and enable PayPal by voice; slots marked literal.original allow special characters and injection; adversarial examples, FGSM and Jacobian attacks, adversarial patches, trojans, data poisoning and model extraction via APIs all work and defenses remain brittle.
AI-written summary · about
chapters (26)
- 0:00Intro
- 1:21Typical Setup at Home
- 4:53Physical Real-World Attacks
- 9:43Splicing Demo 1
- 10:14Splicing Demo 2
- 10:40Future Attacks 1
- 11:49Future Attacks 2
- 12:47Attacking Al Assistant Business Logic
- 14:30Architecture
- 15:06Attackable Slots
- 19:14Neural Networks and the Brain
- 24:15Techniques for Attacking Neural Networks
- 25:05What Can You Attack with Adversarial Examples?
- 25:43Why Do Adversarial Masks work?
- 27:01Adversarial Result
- 27:55Adversarial Input Generation Techniques
- 28:27White Box Adversarial Attack • Techniques
- 30:54White Box Adversarial Attack Techniques
- 31:49Black Box Adversarial Attack
- 33:22Adversarial Patches
- 36:28Defending against adversarial samples
- 38:30Trojaning neural networks
- 42:29Defending against trojans
- 44:43Model Data Extraction
- 46:05Training Data Extraction
- 46:52Summary
related talks
25mHey Google, Activate Spyware! When Google Assistant Uses a Vulnerability as a Feature
57mAI Applications Security Puzzle
20mSecurity in the Age of AI: Response and Strategy
13mMy Voice Is Your Command: The Perils of Smart Voice Assistants
55mMore Tales from the Dark Side: How AI is the Bad Guys New Friend
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
28mHey Google, Activate Spyware! – When Google Assistant Uses a Vulnerability as a Feature
37mFundamentals of AI safety and security