Paper 2026/362

Janus-FHE: Reducing Microarchitectural Leakage in GPU-Based Homomorphic Encryption

Abstract

Homomorphic Encryption (HE) enables secure cloud computing through computations on encrypted data, but the physical execution of HE workloads on shared GPUs can still expose relevant metadata through microarchitectural behavior. We present a case study of BFV relinearization in HEonGPU, a state-of-the-art GPU HE library, and show that repeated executions exhibit structured, reproducible variation in shared-L2 activity across relinearization-key contexts. Our analysis further shows that the regular control-flow structure observed in the optimized HEonGPU binary is compiler-dependent rather than enforced by the source implementation. Together, these observations motivate an HE design in which execution regularity is established explicitly rather than obtained as a consequence of compilation. We present JANUS-FHE, a GPU framework for BFV multiplication and relinearization that restructures the computation around fixed-width large-integer arithmetic. JANUS reconstructs ciphertext coefficients from the RNS representation, uses Kronecker substitution to express polynomial multiplication as large-integer multiplication, and evaluates the resulting convolution through a Stockham-organized Discrete Galois Transform (DGT). JANUS carries this fixed execution structure through product recovery, BFV scale-and-round, and method-I relinearization, using fixed-width computation and branchless masked arithmetic throughout. This formulation makes the control flow, memory-access schedule, and execution geometry depend on public parameters rather than the ciphertext or evaluation-key values. Under the same shared-L2 methodology used for HEonGPU, JANUS exhibits fewer recurring key-conditioned differences, while SASS-level analysis confirms that its intended regular execution structure is preserved after compilation. At \(N=8192\), JANUS completes the full multiplication-and-relinearization path in \(0.350\) ms, compared with \(0.153\) ms for HEonGPU, representing a \(2.29\times\) latency overhead for JANUS's parameter-determined execution structure.