Optimization Theory for Machine Learning
At the foundations of machine learning lie a number of interconnected discrete optimization problems, such as optimal point labeling and subset selection. As time passes, the need to address ever more complex variations on these problems arises. We address this need with a study of the discrete optimization foundations of machine learning, with an emphasis on general-purpose approximation algorithms for high-dimensional discrete optimization.
Objectives:
- Develop & refine the underlying optimization theory that drives learning systems, especially in discrete contexts.
Selected Publications:
- Alan Kuhnle, J David Smith, Victoria Crawford, and My T. Thai. “Fast Maximization of Non-Submodular, Monotonic Functions on the Integer Lattice,” in Proceedings of ICML, 2018
- Minh N. Vu, Truc Nguyen, and My T. Thai. “NeuCEPT: Learn Neural Networks’ Mechanism via Critical Neurons with Precision Guarantee,” in ICDM, 2022
- J David Smith, Alan Kuhnle, and My T. Thai. “An Approximately Optimal Bot for Non-Submodular Social Reconnaissance,” in Proceedings of HyperText, 2018
Privacy & Security in Learning Systems
The ever-growing deployment of machine learning models in industrial and health contexts raises critical privacy and security concerns. These models are built on personal data (e.g. clinical records, images, and user profiles). Our work on this subject focuses on extending ideas in differential privacy to deep neural networks to secure these models in contexts with sensitive data.
Objectives:
- Identify privacy vulnerabilities in existing machine learning systems, particularly those that may leak sensitive information to third parties.
- Develop provably-robust privacy-preserving machine learning systems.
Selected Publications:
- Truc Nguyen & Phung Lai, Nhat Hai Phan, and My T. Thai. “XRand: Differentially Private Defense against Explanation-Guided Attacks,” in AAAI, 2023.
- Truc Nguyen, Phung Lai, Khang Tran, Nhat Hai Phan, and My T. Thai. “Active Membership Inference Attack under Local Differential Privacy in Federated Learning”, in AISTATS, 2023.
- Nhat Hai Phan & Minh Vu & Yang Liu, Ruoming Jin, Dejing Dou, Xintao Wu, and My T. Thai. “Heterogeneous Gaussian Mechanism: Preserving Differential Privacy in Deep Learning with Provable Robustness,” in IJCAI, 2019