Yvonne Zhou - PhD Student

About Me

I am a PhD student at University of Maryland, advised by Professor Dana Dachman-Soled. My research focuses on cryptography.

Research Interests

My research interests are in cryptography. I am particularly focused on developing secure and private machine learning models using advanced techniques like Fully Homomorphic Encryption (FHE) and Differential Privacy (DP). My work also extends to practical security challenges, including building novel secure authentication systems and exploring the security of XR (AR/VR) environments.

Projects

Differentially Private Machine Learning Training under Fully Homomorphic Encryption

Key Concepts: Fully Homomorphic Encryption (FHE), Differential Privacy, Secure Machine Learning, Optimization Theory

Secure and Differentially Private Training under Fully Homomorphic Encryption

Machine learning systems are increasingly trained on sensitive data such as medical records, financial transactions, and government datasets. Protecting privacy during training remains a major challenge, particularly when computation is outsourced to cloud infrastructure. Existing privacy-preserving approaches are often computationally expensive, limiting their use in practice.

This project develops a framework for privacy-preserving machine learning under Fully Homomorphic Encryption (FHE) and Differential Privacy (DP). Training data remains encrypted throughout the entire learning process, ensuring that cloud servers never gain access to the underlying sensitive information. At the same time, differential privacy introduces carefully calibrated randomness during training to prevent the released model from revealing information about any individual data record.

Beyond system design, we provide the first theoretical convergence analysis for machine learning training under encrypted and differentially private computation, demonstrating that accurate learning remains possible despite the approximations and privacy-preserving noise required by these techniques. We further introduce a novel training algorithm that eliminates costly operations, per-sample clipping, used in prior DP-ML methods, significantly improving computational efficiency and scalability.

Experimental results show that our approach achieves model accuracy comparable to existing privacy-preserving training methods while reducing encrypted training costs by several times. By combining strong privacy guarantees with practical performance, this work helps bring secure and trustworthy machine learning closer to deployment in sensitive domains such as healthcare, finance, and government services.

[ See Publication (ICML 2026) ] [ GitHub Repo ]

DP-ML Training on DP Synthetic Data

Key Concepts: Differential Privacy, Machine Learning, Synthetic Data

Conceptual Diagram: Private Data to Public ML Model with DP Synthetic Data

This project develops a preprocessing-based framework for Differentially Private Machine Learning (Pre-DPML), where models are trained on synthetic data that safely replicates key statistical properties of sensitive datasets. Instead of adding noise during model training, we adopt a select–measure–generate workflow: (1) select attribute sets as marginal queries, (2) privately measure their statistics by adding Gaussian noise under (ε, δ)-DP, and (3) generate synthetic data that matches these noisy marginals.

The resulting DP synthetic data enables training of any ML model with privacy guarantees by design, without modifying the learning algorithm. We theoretically bound the excess empirical risk between models trained on real and DP synthetic data and empirically show less than 1% accuracy loss on multiple public datasets. This approach offers strong privacy protection, reusable synthetic data, and high model utility— a practical alternative to traditional training-based DP methods such as DP-SGD and PATE.

[ See Publication (ICML 2024) ] [ GitHub Repo ]

Virtual Login in XR

Technologies: Unity, MongoDB, XR Interaction Toolkit, SHA-256, Raycasting

Virtual Login in XR reimagines user authentication by bringing the login experience into immersive extended reality (XR) environments. Instead of typing traditional passwords, users interact with virtual objects to create unique “virtual passwords.” The project introduces two innovative XR scenes — Diagon Alley and Memory Palace — that leverage spatial memory and object manipulation for secure and engaging authentication.

In Diagon Alley, inspired by the Harry Potter universe, users select specific bricks on a wall to set or verify their password. In Memory Palace, users authenticate by performing distinct movements (rotate, move, lift) on virtual objects in an apartment scene. Each interaction sequence is transformed into a string and securely hashed using SHA-256 before being stored, ensuring confidentiality and protection from reverse-engineering attacks.

By combining principles from cognitive psychology and cryptography, this project enhances memorability, reduces cognitive load, and demonstrates a novel, human-centered approach to authentication in immersive computing environments.

[ Diagon Alley Demo ] [ Memory Palace Demo ]

Publications

Bounding the Excess Risk for Linear Models Trained on Marginal-Preserving, Differentially-Private, Synthetic Data

Yvonne Zhou, Mingyu Liang, Ivan Brugere, Danial Dervovic, Antigoni Polychroniadou, Min Wu, Dana Dachman-Soled

In *Proceedings of the 41st International Conference on Machine Learning (ICML), 2024*. [ PDF ] [ arXiv ] [ Code ]

Revisiting ML Training under Fully Homomorphic Encryption: Convergence Guarantees, Differential Privacy, and Efficient Algorithms

Yvonne Zhou, Mingyu Liang, Ivan Brugere, Danial Dervovic, Yue Guo, Antigoni Polychroniadou, Min Wu, Dana Dachman-Soled

In *Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026*. [ PDF ] [ arXiv ] [ Code ]