Owned By Statistics: How Kubeflow & MLOps Can Help Secure...
summary
Shows concrete attacks and defenses: builds a toy circle-vs-square failure and links it to real failures like a snow detector for huskies, stickers that flip stop signs and a 3D turtle classified as a rifle. Uses Kubeflow pipelines + Kale + Spark for CI/CD, metadata and reproducible training; recommends red teams and feeding production data back. Quantifies extraction risks: ~2,000 queries to clone a logistic regression, ~5,000 for ~99% accuracy, and one million translation queries for <$2k. Defenses include canaries, differential privacy and continuous retraining.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:09Microsoft ML breakthroughs
- 6:01MLOps Benefits
- 8:14Circle Detector
- 12:01Detecting NSFW Content
- 12:47Adversarial Inference
- 14:38Tools to Defend
- 18:40Motivation
- 21:41Language Models
- 22:39The Attack
- 24:57Hidden Data Leakage Examples
- 25:59Secret Memorization
related talks
30mOwned by Statistics: Using Kubeflow to Defend vs Attacks on Your ML Models
40mMLSecOps with Automated Online and Offline ML Model Evaluations on Kubernetes
13mMLOps at Scale
20mML Ops design patterns with Kubeflow Pipelines
26mRunning Secure AI/ML Experiment in a Unified MLOPS-DevSecOps Pipeline
29mTeam Teardown: MLOps in the Cloud for IoT
51mMachine Learning using Kubeflow and Kubernetes
43mDefending Against Adversarial Model Attacks using Kubeflow