Practical Defenses Against Adversarial Machine Learning

Explore practical defenses against adversarial machine learning in this 31-minute Black Hat conference talk. Delve into real-world attacks on various machine learning systems, including recommendation engines, algorithmic trading platforms, email filtering, facial recognition, and malware classification. Gain insights from research conducted over a year, moving beyond simplistic gradient-based comparisons to understand the actual attack landscape and assess risks accurately. Learn about calibrated mitigations for real threats, covering topics such as bad inputs, model leakage, block lists, multiple signals, and raw statistics. Discover recommendations for defense strategies, open-source projects, partial homomorphic encryption, federated learning, and handling incomplete data. Examine vendor examples, compare deep fakes to defects, and discuss the implications of larger models in the context of adversarial machine learning.

Intro
Who am I
Research vs Deployment
Bad Inputs
Email Filtering
Transportation Prediction
Recommendation Engines
Trading Bots
Model Leakage
Block Lists
Multiple Signals
Raw Statistics
Conclusion
Recommendations
QA
Open Source Projects
Partial Homomorphic
Federated Learning
Incomplete Data
Contact
Vendor Examples
Deep Fakes vs Defects
Larger Models
Deep Fakes
Outro