Machine Learning Models in a Dynamic Environment
summary
Users adapt to models and break assumptions: historical examples (SEO gaming, email spam) show adversarial cat‑and‑mouse dynamics, LLMs trained on the web create a feedback loop when model output is re‑published, and crawled training data can be poisoned via open repos and npm packages. Practical defenses are concrete: validate training inputs like SQL injection, monitor covariate shift and anomalous user cohorts with automated detectors, share signals across platforms as banks do for fraud, and run the live Mentimeter barcode exercise and note the Wordfreak shutdown as cautionary evidence.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:12Open source and machine learning background
- 2:38Search engines history
- 4:52Attention competition
- 5:56Email spam detection
- 7:44Fraud detection challenges
- 8:26Social media ranking
- 11:01Interactive ranking game
- 14:21Social media attacks
- 16:28Sock puppet detection
- 18:09Malicious behavior detection
- 19:36Platform cooperation
- 20:59Input validation
- 22:45Adversarial machine learning and covariate shift
- 23:29LLM feedback loops
- 24:52LLM side effects
- 27:06Adapting to user behavior
- 28:35Tech and society
- 30:21Adversarial patterns shift
- 31:45AI education
- 32:26Open floor for questions and discussion
related talks
20mMachine Learning Attacks – A New Era of Security Epidemic
28mSecurity of Machine Learning
39mChanging Realities - The Final Frontier for Machine Learning?
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
18mReliable Maintenance of Machine Learning Models
53mSecurity of Machine Learning Systems
24mEngaging new and diverse audiences with ML
01h 04mCan you trust your (large language) model?