Paper 2025/1775

Homomorphic Encryption Methods Applied to Cloud Computing: A Practical Architecture for Elastic, Verifiable Confidential Compute

Abstract

Cloud computing has matured into the default substrate for data processing, yet confidentiality demands of- ten force a hard trade-off between the utility of outsourced computation and the privacy of sensitive inputs. Homomorphic encryption (HE)[1] promises to dissolve that trade-off by enabling computation directly on ciphertexts, returning encrypted results that only the data owner can decrypt. Despite remarkable progress in fully homomorphic encryption (FHE) and leveled variants suitable for bounded-depth circuits, deploying HE at cloud scale remains challenging. The cost of ciphertext arithmetic is orders of magnitude higher than plaintext compute; noise growth and rescaling impose algorithmic discipline; vectorization and rotation keys complicate program layout; and the lack of verifiability in bare HE obliges trust in a correct-but-curious cloud[?]. This paper develops a system perspective on how to apply modern HE in cloud environments without reducing it to a boutique feature. We introduce a reference architecture that marries approximate-arithmetic HE for analytics with exact- arithmetic HE for integrity-critical operations, composes HE with succinct proofs for verifiability, and integrates a cost model into the scheduler so that elastically provisioned serverless workers can meet latency objectives under price constraints. The design begins with a compiler that lowers dataflow graphs to operator sequences parameterized by multiplicative depth L and rotation sets; it then chooses schemes and parameters—CKKS for floating-point style analytics and signal processing, BFV/BGV for integer operations and counters, TFHE-style bootstrapping for comparisons—that minimize the total time-to-result under explicit error and security budgets. A cryptographic key service supports threshold issuance and rotation-key escrow without learning plaintexts, while a storage layer packs columns into ciphertext SIMD lanes to amortize cost across tenants. For verifiability, we attach homomorphic message authentication tags to intermediate ciphertexts and wrap end-to-end executions in succinct non-interactive proofs specialized to the bilinear equations that certify correct key switching, rescaling, and boot- strapping. Analytically, we characterize latency by a linear model in the counts of core homomorphic primitives and show how to saturate GPUs or vector units with batched number-theoretic transforms to bend throughput toward practical regimes. Under realistic traces of analytic queries and encrypted inference, the architecture achieves sub-second P95 for circuits of depth six to eight with one or two bootstraps, while sustaining 128-bit security under RLWE. By treating HE not as an exotic afterthought but as a first-class cloud programming and scheduling primitive, the proposed approach demonstrates a path to confidential-by- default services in which the cloud never sees data in the clear yet remains efficient, elastic, and auditable.