FPGA Acceleration of Homomorphic Encryption-Assisted Federated Learning

Cloud computing is becoming increasingly popular and widely used in daily services. However, the traditional computation on the cloud involves decryption and processing plaintext on the server side, raising privacy and security concerns. Homomorphic encryption enables computations on encrypted data without decryption, but is significantly slower. Accelerating these computations through hardware and arithmetic optimizations, like FPGA usage, is essential. Previous works focused on implementing building blocks for constructing a homomorphic encryption system, but do not provide a full framework to perform full operations. In this work, we go in a different direction: We design a client-server framework with a communication channel and an FPGA-based acceleration for the Fan-Vercauteren algorithm, as it has several open-source implementations and can be easily deployed. Using the framework, we evaluate the end-to-end performance of the algorithm. The client employs software for encryption, decryption and key generation, while the server utilizes FPGA hardware to accelerate the computation. Furthermore, we evaluate the timing needed and compare it with executing a full software version of the system. We further extend and adapt the framework to use additive homomorphic encryption in the federated learning use case, where the cloud server is responsible for aggregation of the clients’ models. We evaluate our extended framework and record the aggregation time and resource utilization of the XI-based Paillier adder.