ฮจ
PSI
Private
& Safe
Intelligence
We study the theoretical foundations of AI and develop privacy- and safety-preserving algorithms. Our research covers three areas: understanding AI (generative modeling, generalization, and memorization), privacy & safety (differential privacy, machine unlearning, and attack robustness), and industrial applications (time series, finance, and manufacturing).
1 ๐Ÿง 

Understanding of AI

We study the theoretical foundations of modern AI. Our goal is to understand how AI works.

๐ŸŽจ

Generative Modeling

Developing the fundamental algorithms of generative AI, such as diffusion models.

๐Ÿ“Š

Generalization

Analyzing loss landscape geometry to understand when and why deep networks generalize.

โšก

Memorization

Investigating why memorization happens in AI and how it is related to generalization.

2 ๐Ÿ”’

Privacy & Safety Issues of AI

We develop privacy- and safety-preserving machine learning algorithms. Our work spans formal privacy guarantees via differential privacy, machine unlearning for generative AI, and defenses against AI attacks.

๐Ÿ›ก๏ธ

Differential Privacy

Optimizing the utility-privacy tradeoff for ML models trained with DP-SGD, including private SVMs and synthetic data generation.

๐Ÿ—‘๏ธ

Machine Unlearning

Efficiently removing specific data or concepts from trained models (including diffusion models) without full retraining.

โš”๏ธ

Attack Robustness

Understanding robustness and building resilience against attacks such as adversarial attacks and membership inference.

3 ๐Ÿ“ˆ

Industrial Applications

We apply trustworthy AI techniques to real-world industrial challenges.

๐Ÿ“‰

Time Series

Robust forecasting and synthetic data generation for noisy, scarce sequential data.

๐Ÿ’น

Finance

Asset price modeling with generative diffusion models and applications of AI in financial fields.

๐Ÿญ

Manufacturing

Generative AI for industrial domain and privacy concerns for industrial data.