SparkFHE: Distributed Dataflow Framework with Fully Homomorphic Encryption

We propose a new framework, which aims to enable large-scale privacy-preserving machine learning (PPML) in the cloud. In this extended abstract, we discuss the integration of Apache Spark and fully homomorphic encryption (FHE), thus the name SparkFHE. The SparkFHE framework enables Spark to perform computations on encrypted data without requiring the secret key hence preserving user data privacy, and scales up homomorphic algorithms for larger datasets through efficient cluster programming. We present the architecture design, programming abstractions, and mappings of applications such as private set intersection and logistic regression, into the dataflow model of Spark. Furthermore, we discuss preliminary results to validate our designs.