From Bottlenecks to Breakthroughs: Accelerating MPC for Secure ML
Abstract: Secure multi-party computation (MPC) holds the promise of enabling privacy-preserving machine learning across data and model silos—but in practice, performance, scale and useability bottlenecks have limited real-world adoption. In this talk, I will discuss how recent advances in function secret sharing (FSS) are transforming these bottlenecks into breakthroughs, pushing MPC for secure ML from theory to high-performance reality. ORCA combines novel FSS-based protocol designs with GPU acceleration to speed up secure training and inference – achieving sub-second ImageNet inference. SIGMA brings secure transformer inference into the realm of practicality, introducing new FSS-based protocols for core ML functions and enabling the first secure execution of GPT-class models, including LLaMA2-13B in under a minute. Finally, I will discuss SHARK, the first FSS-based system for actively secure ML inference that outperforms prior state-of-the-art protocols by two-three orders of magnitude.
Date and Time
Location
Hosts
Registration
Add Event to Calendar
Speakers
Divya Gupta of Microsoft Research India
From Bottlenecks to Breakthroughs: Accelerating MPC for Secure ML
Abstract: Secure multi-party computation (MPC) holds the promise of enabling privacy-preserving machine learning across data and model silos—but in practice, performance, scale and useability bottlenecks have limited real-world adoption. In this talk, I will discuss how recent advances in function secret sharing (FSS) are transforming these bottlenecks into breakthroughs, pushing MPC for secure ML from theory to high-performance reality. ORCA combines novel FSS-based protocol designs with GPU acceleration to speed up secure training and inference – achieving sub-second ImageNet inference. SIGMA brings secure transformer inference into the realm of practicality, introducing new FSS-based protocols for core ML functions and enabling the first secure execution of GPT-class models, including LLaMA2-13B in under a minute. Finally, I will discuss SHARK, the first FSS-based system for actively secure ML inference that outperforms prior state-of-the-art protocols by two-three orders of magnitude.
Biography:
Divya Gupta is a Principal Researcher at Microsoft Research India. Her research interests lie in cryptography, particularly secure computation, blockchains, and, more recently, AI security.
Before joining Microsoft Research, she was a postdoctoral researcher at UC Berkeley, hosted by Sanjam Garg. She completed her Ph.D. at the University of California, Los Angeles, under the guidance of Amit Sahai. Her dissertation was recognized with the Dissertation Fellowship and the Dimitris N. Chorafas Dissertation Award, presented for outstanding work in engineering sciences, medicine, and the natural sciences. Prior to that, she earned her Bachelor’s and Master’s degrees in Computer Science from the Indian Institute of Technology Delhi, where she was recently honored with the Graduates of the Last Decade (GOLD) Award by her alma mater.
She has published extensively in leading cryptography and security venues, including CRYPTO, EUROCRYPT, TCC, IEEE S&P, CCS, and USENIX Security, and has served on the technical program committees of many of these conferences. More recently, her research has helped shape the security architecture for Microsoft’s Copilot Tuning product. She also holds several U.S. patents.