Making & Breaking Machine Learning Anomaly Detectors in Real Life

on YouTube ↗01h 00m

" summary

Live demos show a practical model-poisoning attack that slowly injects crafted web-log traffic into PCA-based detectors; implementation uses PCA and a robust PCA variant on Apache access logs (10-week training, ~100k records). Injecting 30–50% poisoned samples via a “boiling frog” slow ramp shifts decision boundaries and yields about 38% evasion success against robust PCA in these experiments, and naive PCA weakens even more — scikit-learn style implementations need feature care and adversary-aware evaluation.

AI-written summary · about

" chapters (45)

1. 0:00Introduction
2. 1:14Agenda
3. 1:40What is Machine Learning
4. 2:06What is Anomaly Detection
5. 2:37Intrusion Detection
6. 7:05Machine Learning Libraries
7. 8:03Why is Machine Learning Good
8. 10:24Machine Learning is Not the Right Answer
9. 10:40Why Threshold Static Rule Base Heuristics are Good
10. 11:36Machine Learning Applications
11. 13:18Sentiment Analysis
12. 15:11Expectations
13. 15:41Research
14. 17:10Machine Learning
15. 21:47Lack of Training Data
16. 22:45Semantic Gap
17. 24:54Evaluation
18. 26:58Adversary Impact
19. 28:05Anomaly Detection System Fails
20. 29:32Machine Learning Anomaly Detection Hopeless
21. 30:42Anomaly Detection Example
22. 32:55Common Machine Learning Techniques
23. 35:37Selecting Features
24. 37:47Principal Component Analysis
25. 40:23Avoiding Common Pitfalls
26. 41:12Semantic Gaps
27. 41:35Evaluation Techniques
28. 42:10True Positives
29. 44:22Attack a Model
30. 45:15Chef
31. 46:31Asset X
32. 47:17Boiling Frog Attack
33. 47:44How to Avoid Detection
34. 48:09How to Defend Against Detection
35. 48:44Decision Boundary Rage Ratio
36. 49:15Can ML Be Secure
37. 49:43PCA vs Median
38. 50:30Simulations
39. 51:33Linear Regression Model
40. 52:56Detection Logic
41. 53:53Detection Graph
42. 55:21Evasion Success
43. 56:52What Next
44. 57:22Meetup Group
45. 58:23Anomaly Data

" related talks

• 20mMachine Learning Attacks – A New Era of Security EpidemicAshima Purohit · Developer Summit 2020
• 33mPoisoning Web-Scale Training Datasets is PracticalWill Pearce · Black Hat USA 2023
• 28mSecurity of Machine LearningJelena Milosevic · Data Science Conference 5.0
• 46mA Short Introduction to Hacking Artificial IntelligenceProf. Jerzy Surma · Oh My H@ck 2020
• 30mMachine Learning Techniques for Class Imbalances & AdversariesBrendan Herger · PyData DC 2016
• 41mMachine Learning Models in a Dynamic EnvironmentIsabel Drost-Fromm · PyCon DE & PyData 2025
• 39mDetecting Website Intrusion and Account Compromise with Machine LearningRobin Franklin Guha · RSA Conference 2024
• 46mAI in a Minefield: Learning from Poisoned DataJohnathan Azaria · Big Data Conference Europe 2022