Skip to content

SWIFT: Shallow and SIMD-Aware CKKS Functional Bootstrapping for Low-Latency

Year: 2026

type: technique collection: schemes status: auto

Links: ePrint 2026/1163

Summary

Auto-extracted summary (unreviewed)

SWIFT is a CKKS functional-bootstrapping method for moderate batch sizes that packs integer multiples of each input into unused SIMD slots so that exponential bootstrapping produces the powers exp(2piilx) in parallel, replacing homomorphic power computation and thereby lowering multiplicative depth, ciphertext modulus and ring degree; a convex-optimization procedure also finds trigonometric approximations with small coefficient norm for singular functions such as 1/sqrt(x). At log N = 15 and 12-bit precision it evaluates ReLU and 1/sqrt(x) 3.94x and 2.86x faster than CKKS baselines for 512 inputs, reaches 4.48x and 14.68x for 8-to-8 and 12-to-12 LUTs at batch 32, and beats latency estimates of DM/CGGI-based methods at batches as small as 32.

Abstract

The CKKS homomorphic encryption scheme can achieve high throughput for large batches, typically of tens of thousands of inputs, whereas DM/CGGI schemes offer low latency for a single input. In existing CKKS approaches, the multiplicative depth required for nonlinear function evaluation demands a large ciphertext modulus, which forces a large ring degree regardless of the input size. Practical applications, such as encrypted MLP inference, may require only a few hundred function evaluations at a time. For these moderate batch sizes, existing CKKS-based methods requiring large multiplicative depth incur high latency without fully benefiting from their high throughput. We present SWIFT, a CKKS functional-bootstrapping method that reduces latency by using multiple slots to lower multiplicative depth. For input x, SWIFT packs its integer multiples ℓx into unused slots. Exponential bootstrapping then generates the powers exp(2πiℓx) = exp(2πix)ℓ= αℓin parallel. Compared to previous approaches that compute these powers through homomorphic multiplications, SWIFT directly generates those powers and reduces the required multiplicative depth and ciphertext modulus, allowing a smaller ring degree and lower latency. We also introduce a trigonometric approximation method for singular functions such as 1/√x, whose derivative is unbounded near zero. We observe that the previous Fourier extension method of Bian et al. (EU- ROCRYPT’26) can produce extremely large coefficient norms for these functions when focusing solely on reducing the polynomial degree. Instead, we solve an optimization problem to find a trigonometric polynomial that minimizes the coefficient norm while satisfying the required approximation accuracy for the target function. The resulting trigonometric polynomial can be evaluated efficiently and with numerical stability using SWIFT. Our implementation evaluates ReLU on [−1, 1] and 1/√x on [0.0005, 1] with 12-bit absolute and relative precision, respectively, at ring degree log N = 15. For batches of 512 inputs, SWIFT achieves speedups of 3.94× and 2.86× over the corresponding CKKS baselines for ReLU and 1/√x, respectively. Compared with latency estimates of DM/CGGIbased methods, SWIFT achieves lower latency for 8-to-8 LUT and 12to-12 LUT at batch sizes as small as 32 inputs.
Source: pdf

Extracted details

Field Value Source Confidence
contribution slot packing of integer multiples of inputs to generate exponential powers in parallel during bootstrapping llm 0.85
contribution tensor-contraction polynomial evaluation reducing multiplicative depth and key-switching count versus Paterson-Stockmeyer llm 0.85
contribution numerical error bounds linear in the trigonometric coefficient norm llm 0.85
contribution convex optimization for minimum-coefficient-norm trigonometric approximations of singular functions (1/x, 1/sqrt(x)) llm 0.85
contribution implementation and comparison on LUT, sigmoid, tanh, ReLU, 1/x and 1/sqrt(x) tasks llm 0.85
schemes CKKS llm 0.85
libraries TFHE-rs llm 0.85
libraries lattice-estimator llm 0.85
tasks functional bootstrapping / look-up table evaluation llm 0.85
tasks activation functions (ReLU, sigmoid, tanh) llm 0.85
tasks inverse and inverse square root evaluation llm 0.85
security level 128-bit llm 0.85
interaction non-interactive FHE llm 0.85
limitations choice of tensor factor sizes is not optimized for different polynomial degrees llm 0.85
llm relevance low llm 0.85
key result speed-up over CKKS baseline (ReLU): 3.94x (512 inputs, [-1,1], 12-bit absolute precision, log N = 15) llm 0.85
key result speed-up over CKKS baseline (1/sqrt(x)): 2.86x (3.19 s for 512 inputs) ([0.0005,1], 12-bit relative precision, log N = 15) llm 0.85
key result speed-up over CKKS baseline (8-to-8 LUT): 4.48x (batch size 32) llm 0.85
key result speed-up over CKKS baseline (12-to-12 LUT): 14.68x (batch size 32) llm 0.85
key result latency vs DM/CGGI estimates: lower for 8-to-8 and 12-to-12 LUT (batch sizes as small as 32) llm 0.85

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: CKKS 104 [p.1], TFHE 18 [p.1], BFV 2 [p.6], FHEW 2 [p.29]
  • library: TFHE-rs 3 [p.22], HElib 1 [p.29]
  • model: MLP 3 [p.1], LLaMA 1 [p.30], CNN 1 [p.29]
  • hardware: Intel Xeon 3 [p.23]
  • technique: Bootstrapping 81 [p.1], Lookup table 65 [p.1], SIMD packing 36 [p.1], Rotations 32 [p.6], Functional bootstrapping 17 [p.1], Key switching 17 [p.3], Sigmoid 16 [p.22], ReLU 15 [p.1], Programmable bootstrapping 5 [p.2], Paterson-Stockmeyer 5 [p.2], Relinearization 3 [p.7], Rescaling 3 [p.15], Chebyshev 3 [p.22]
  • assumption: Lattice-based 4 [p.24]
  • topic: Homomorphic encryption 14 [p.1], Benchmark 3 [p.30]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/schemes/s13-2_eprint-2026-1163.pdf schemes 52 962.5 KB 3dd867d7f0e8

Database key: swift-shallow-and-simd-aware-ckks-functional-bootstrapping-for-low · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set swift-shallow-and-simd-aware-ckks-functional-bootstrapping-for-low COLUMN VALUE