Trustworthy Machine Learning: Challenges and Frameworks

Explore the critical aspects of trustworthy machine learning in this 18-minute conference talk from USENIX Enigma 2020. Delve into the expansive attack surface of ML systems, including data poisoning, adversarial examples, and model exploitation. Examine the urgent need for security considerations in ML algorithm design and the opportunity to address these issues before widespread deployment. Learn about a framework for fostering trust in ML algorithms, uncovering the influence of training data on predictions, and identifying potential security and privacy risks. Gain insights into interpreting model behavior and extracting essential data representations for trustworthy machine learning. Cover topics such as safety, privacy, ethical aspects, differential privacy, stochastic gradient descent, and model governance.