Owned By Statistics - How Kubeflow & MLOps Can Help Secure Your ML Workloads

Overview

Explore how Kubeflow and MLOps can enhance the security of machine learning workloads in this 40-minute conference talk by David Aronchick from Microsoft. Delve into the challenges of transitioning machine learning models from development to production, focusing on potential vulnerabilities and degradation risks. Learn about implementing a robust MLOps process using Kubeflow to address common pitfalls in machine learning workflows. Discover techniques for ensuring reproducibility, validation, versioning, tracking, and safe deployment of ML models. Gain insights into the future direction of MLOps and its potential to accelerate development while mitigating risks. Examine various types of attacks on ML models, including adversarial examples and data poisoning, and explore strategies to defend against them. Understand the importance of building efficient MLOps pipelines to continuously improve model performance and security. Discuss the reality of ML security threats and the necessity of proactive measures in safeguarding your machine learning workloads.

Syllabus

Introduction
Machine Learning at Microsoft
ML in every product at Microsoft
ML in the average enterprise
Data scientist
Building a model
Rolling it out
Security
Three types of attacks
Advanced models
Snow detection
Stop sign detection
Face recognition
Defend against adversaries
Build an MLOps pipeline
Modular components
Pipeline example
Another attack vector
Malicious users
Two types of attacks
Distillation attack
Accuracy
GoogleBERT
Continuous Improvement
Build Efficient Pipelines
Take Your Models
Hidden Data
Recommendations
Network Graph
Map Leakage
Example
How to prevent this
Injections
Leaks
Summary
The Reality
You will be attacked
Conclusion
Questions
Reprocessing ML Pipeline Predictions
MLOps vs Continuous Machine Learning
Regulation of ML
Mitigating Leaky Data