Distributed Edge-Adaptive Parallel Binary Fully Homomorphic Encryption Framework

Abstract

The recent growth in the privacy-sensitive artificial intelligence of distributed cloud-edge systems has accelerated the necessity of the implementation of efficient and thermally feasible encrypted inference engines. Fully Homomorphic Encryption (FHE) makes it possible to perform computation on encrypted data, and its expensive nature of computation, ciphertext expansion, and communication cost are serious barriers to scaling in heterogeneous settings. This paper presents Distributed Edge-Adaptive Parallel Binary Fully Homomorphic Encryption Framework (DEAP-BFHE) as a resourceful distributed inference framework that incorporates parallel binary FHE processing, AI-based parameter tuning at run time and workload-sensitive ciphertext migration. The framework is a dynamically optimizing encryption depth, batching size and resource allocation framework on real-time monitoring of Thermal Power Density (TPD), Ciphertext Communication Overhead (CCO), Scalability Efficiency (SE) and Resource Utilization Ratio (RUR). The CIFAR-100 dataset was used to evaluate performance on Multi2Sim with McPAT integrated simulator and compared with CHET, nGraph-HE2, and GAZELLE at the same setting. The experimental findings show that DEAP-BFHE can minimize the Thermal Power Density by up to 31.7%, Ciphertext Communication Overhead by 31.2%, Scalability Efficiency by 29.2%, and Resource Utilization Ratio by 19.2% in comparison with CHET using 32 nodes. These results support the fact that adaptive encryption parameter optimization and smart ciphertext migration can contribute greatly to scalability, energy, and communication efficiency. The proposed framework gives a strong basis to AI deployment with sustainable and large-scale use, preserving privacy in next-generation heterogeneous cloud-edge systems.