01 May 2020
Regency Ballroom IV
Deep Learning and Extracting Insights from Encrypted Data with Darknet: Lessons Learnt and Challenges Ahead (G31b)
Learning from encrypted data can address some of the primary concerns related to privacy, propriety, and legality of sharing sensitive data and potentially enable federated learning to gain insights from otherwise isolated collections of non-shareable data from different stakeholders. In this work, we explore Fully Homomorphic Encryption (FHE) as a promising technique to enable analytics while providing strict guarantees against information leakage. We propose to explore FHE technique for both inference and training. We measure the computational effort needed to apply each atomic operator and related performance metrics. These estimates of the atomic operators are in turn used to forecast the overall computational effort needed for secure inference/training and validate the estimates for a small dataset, MNIST in various configuration settings. We believe that lessons learned from our experiments will constitute valuable insights for ML and Crypto community to join forces to solve the fundamental scientific challenges underlying learning from encrypted data together.