Short Wave How Hackers Could Fool Artificial Intelligence
Sep 21, 2020
Artificial intelligence may not be as foolproof as we imagine. Researchers reveal how simple tweaks can lead AI systems astray, especially in critical areas like self-driving cars. The discussion highlights the urgent need for stronger security and resilience in AI technologies. Plus, there's a light-hearted moment that links these serious risks to a broader conversation about financial literacy. Tune in for insights into the unpredictable world of adversarial AI!
AI Snips
Chapters
Transcript
Episode notes
AI Vulnerability
- AI decision-making is vulnerable due to its reliance on data points and shortcuts, unlike humans who consider context.
- This vulnerability allows hackers to manipulate specific data points, tricking AI into misidentification.
Disco Deception
- An example involves training AI to identify disco music by focusing on elements like beats per minute and instruments.
- By subtly adding an oboe, undetectable to humans but noticeable to AI, the system misclassifies the music.
Facial Recognition Fooled
- Researchers tricked facial recognition AI by using large, colorful glasses.
- The AI misidentified the subject due to the glasses.