[Seminar] Privacy-Preserving Deep Learning
Monday, February 21, 2022
11:00 am - 12:00 pm
Senior Research Scientist
Samsung Research America
*Microsoft 365 @cougarnet.uh.edu authentication required to join via Zoom
Recent years have witnessed the rapid development of artificial intelligence and deep learning. Applying deep learning models to problems that involve biomedical, financial, military, and other sensitive data requires protecting data privacy. Cryptography protocols, e.g., Fully Homomorphic Encryption (FHE) and Multi-Party Computation (MPC), are becoming increasingly popular methods to design Privacy-Preserving Deep Learning (PPDL). However, practical PPDL with desired accuracy and performance cannot be provided by the state-of-the-art works, i.e., existing PPDL suffers from prohibitively computational or communicational overhead. To design efficient PPDL, I proposed crypto-friendly deep learning methods, including deep activation replacement, homomorphic frequency-domain convolution, FHE-friendly neural network architecture search, and novel cryptography protocols that are adaptive for deep learning. Extensive experiments on real-world medical datasets, e.g., diabetic retinopathy, skin cancer, genome base-calling, show that the proposed methods provide a practical, fast, and accurate PPDL.
About the Speaker
Qian Lou, is a senior research scientist at Samsung Research America. Prior to that, he obtained his master’s and Ph.D. degrees from Indiana University. He has received several awards, like Samsung Research Q4 best paper award and the best paper nomination at the PACT conference. More than 20 papers are published at top-tier conferences. He worked as an organization program committee member for the FHE.org conference and serves as a program committee member at multiple conferences, such as NeurIPS, ICML, AAAI, ICLR, GLSVLSI. He is also a member of IEEE and ACM.