Owned by Statistics: Using Kubeflow to Defend vs Attacks on Your ML Models
Talkat KubeCon + CloudNativeCon North America 2020 Virtual · on Wednesday, 2020-11-18· at 17:45· Intrado Virtual Event Platform,·
summary
Shows how a cloud-native MLops pipeline built with Kubeflow Pipelines plus Kale (notebook→pipeline), Katib for hyperparameter search, Rok snapshotting and ML Metadata and KFServing closes the loop: ingest → train → hyperparameter sweep → serve → monitor. Demonstrates attacks: adversarial stickers flip stop signs, 3D-printed objects spoof classifiers, face recognition mislabels 28 congresspeople, model extraction reproduces models with ~5,000 queries and reaches ~72–86% fidelity with far fewer probes. Emphasizes fast retrain, red‑teaming and automated detection as the practical defenses.
AI-written summary · about
chapters (11)
- 0:00Start
- 0:34Microsoft ML breakthroughs
- 3:38Circle Detector
- 5:15The Explanation Reveals Why
- 6:15Adversarial Inference
- 7:28Tools to Defend
- 21:50Language Models
- 22:54The Attack
- 25:33Hidden Data Leakage Examples
- 26:28Secret Memorization
- 29:18Truths You Cannot Avoid
related talks
29mOwned By Statistics: How Kubeflow & MLOps Can Help Secure...
43mDefending Against Adversarial Model Attacks using Kubeflow
21mDefending Against Adversarial Model Attacks Using Kubeflow
40mMLSecOps with Automated Online and Offline ML Model Evaluations on Kubernetes
51mMachine Learning using Kubeflow and Kubernetes
39mOperating Deep Learning Pipelines Anywhere Using Kubeflow
39mDeploy your ML Workloads on Kubernetes with Kubeflow
46mAccelerating Machine Learning DevOps with Kubeflow