Trustworthy Machine Learning

Course Description

This is an advanced topics course on Trustworthy Machine Learning (ML) with a focus on the following contemporary and increasingly important design aspects: robustness of ML models; data privacy of users that is used for learning, and bias and fairness in ML. Recent developments have shown that ML models showing state-of-the-art accuracies can be readily misled by making small (imperceptible) perturbations to the input. As ML based approaches continue to be adopted for decision making in safety critical domains and applications (such as healthcare, military/defense, autonomous vehicles), the issue of robustness is central to ensuring widespread adoption of trustworthy ML/AI. This has led to an explosion of new research in the area of adversarial machine learning which deals with getting an understanding of why neural network-based ML models show such brittle behavior and ultimately incorporate robustness in the ML training pipeline. At the same time, ML models (such as recommendation systems or LLMs) are typically trained on personal data collected from users. Such models have been shown to possess critical vulnerabilities, i.e., it is possible to extract private information of users just by interacting with the model.

Code

ECE 522

Credits

3

Subject

Elec Comp & Sftwr Engr, Sch of