Robots, Manners and Stress
summary
Claim: machine-learning components in robots get stressed and provoke human backlash, so alarms beat full robustness. Demonstrations include adversarial images made by gradient descent that turn a bird into a car, a taboo trap defense that diversifies models per customer, and sponge examples that make NLP or vision systems use orders of magnitude more CPU (one Google Translate case slows by ~5,000×). Proposal: practical systems alarm and fall back—e.g. reduce speed to 20 km/h and request human takeover—rather than try to be invulnerable.
AI-written summary · about
chapters (15)
- 0:00Intro
- 2:16How not to behave
- 3:40When robots and humans mix it (2)
- 4:54Situational awareness
- 6:24Automatic driver assistance systems
- 7:41Adversarial inputs
- 9:16Attack Detection (2)
- 9:58Idea: the Taboo Trap
- 11:47Other alarms
- 12:44Sponge attacks
- 13:26Bad characters
- 14:53Security engineering for ML 101
- 15:57A role of manners?
- 17:42Pulling it all together
- 18:45Future research questions
related talks
52mSome pitfalls in AI
41mThe High Stakes Security of AI
50mSecurity as Part of Responsible AI: At Home or At Odds?
16mOn the ethical challenges raised by robots powered by AI
44mToxic Trust: The Psychology of Automation Bias in Generative AI
44mComputers are Stupid: Protecting "AI" from Itself
06m7 practical tips for safer AI and language models
42mIs your AI plotting to kill you? Why AI and Machine Learning is dangerous