Research
ParkLab develops machine learning methods for privacy-preserving data analysis, synthetic data generation, and safer foundation-model use.
Privacy-preserving machine learning
We study algorithms that enable data analysis without sacrificing individual privacy. Our work uses differential privacy as a rigorous mathematical foundation and aims to improve the trade-off between privacy and accuracy.
Differentially private synthetic data
Synthetic data can make sensitive data easier to share and study, but only when privacy guarantees are carefully designed and evaluated. Our group develops methods for private data generation, data distillation, and private sampling.
Safety in diffusion and foundation models
Recent work studies safe use of diffusion-based foundation models, text diffusion models, safe denoisers, safety-guided flow, and alignment against adversarial jailbreaks.
Bayesian learning and model understanding
We also work on Bayesian machine learning, including efficient inference, approximate Bayesian computation, Bayesian deep learning, and Bayesian perspectives on prompts and foundation models.
Selected support
- CAISI Safety Catalyst: On the Safe Use of Diffusion-Based Foundation Models.
- DSI postdoctoral matching funding: Safe Diffusion Models.
- NSERC Discovery: Privacy-preserving machine learning.
- CIFAR–Amii AI Chair
- Roche and CIFAR support for machine learning in healthcare