Close

Presentation

PRIMER: A Privacy-Preserving Transformer on Encrypted Data
DescriptionNLP has witnessed rapid progress driven by Transformer based models, but applying Transformer models requires attention to data privacy and security. Cloud-based Transformer solutions for such tasks are limited by legal and ethical requirements. In this paper, we are the first to present a privacy-preserving Transformer with high throughput and accuracy, called Primer. Primer enables a data owner to send encrypted data to a server-based Transformer that can be directly applied to encrypted data, thereby preserving data owner's privacy. Our experimental results show Primer achieves SOTA of NLP accuracy and reduces the inference latency by 73.6% ∼ 97.1% over Primer-base.
Event Type
Research Manuscript
TimeTuesday, July 11th2:25pm - 2:40pm PDT
Location3004, 3rd Floor
Topics
AI
Keywords
AI/ML Security/Privacy