MLSecOps with Automated Online and Offline ML Model Evaluations on Kubernetes
summary
Demonstrates a concrete ML security stack: Adversarial Robustness Toolbox (Python) plus AIF360/XAI toolkits integrated into Kubeflow + KServe for online explainers and offline logging. Shows measurable impact: a vanilla MNIST model falls from 87% to 13% under FGSM and recovers to 94% with spatial smoothing. Describes a working pipeline using Knative/Kafka → MySQL ingestion, Prometheus/Grafana monitoring, and an offline evaluator that flags fairness/robustness regressions; project stats: 3000+ stars, 150k+ downloads.
AI-written summary · about
chapters (10)
- 0:00Start
- 11:07Financial Check Fraud
- 12:15Adversarial Threats to Al
- 13:02Real-world Adversarial Exploits
- 13:59Adversarial Threats to Machine Learning
- 20:46Pipeline: Adversarial Robustness Toolbox (ART) Component
- 24:05Online vs Offline model evaluations
- 26:19Inference Service Control Plane
- 28:44ART Explainer Spec
- 29:14ART Online Evaluation Demo Flow
related talks
40mMLSecOps with Automated Online and Offline ML Model Evaluations on Kubernetes
30mOwned by Statistics: Using Kubeflow to Defend vs Attacks on Your ML Models
26mRunning Secure AI/ML Experiment in a Unified MLOPS-DevSecOps Pipeline
29mOwned By Statistics: How Kubeflow & MLOps Can Help Secure...
31mApplied MLOps to Maintain Model Freshness on Kubernetes
21mEnd-to-end secure ML Development
25mSecure ML: Automated Security Best Practices in Machine Learning
42mMLOps Under Attack: Threat Modeling Modern AI Systems