At the age of AI and ML
summary
Shows a concrete pipeline that detects adversarial inputs by capturing layer activations with SageMaker Debugger, comparing training vs inference representations via t-SNE and an MMD two-sample test, and scheduling hourly Model Monitor checks that write violations to S3 and CloudWatch. Uses PGD attacks (epsilon up to 0.5) to flip labels, finds adversarial separation emerges in the ResNet18 penultimate layer, and reports detection rates that hit ~100% during the attack window before dropping.
AI-written summary · about
chapters (12)
- 0:00Chaos engineering with AI
- 1:05Welcome
- 2:10Session agenda
- 3:47ML system failures
- 9:10AI/ML complexity
- 11:20Traditional vs ML testing
- 14:11Adversarial ML threats
- 15:58Detecting adversarial samples
- 23:01Model monitoring with SageMaker
- 26:13SageMaker Debugger
- 28:21Adversarial input detection
- 31:36Resilient ML systems
related talks
31mBuild threat detection models using ML
31mMonitoring machine learning models in production
25mMonitoring AI with AI
21mEnd-to-end secure ML Development
27mWaking the Data Scientist at 2am: Detect...
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
30mTrain ML models faster with better insights using Amazon SageMaker Debugger
38mUsing adversarial samples to robustify your Neural Network Models