Source-linked AI summary
High-speed Implementation of Length-compatible Privacy Amplification in Continuous-variable Quantum Key Distribution
Xiangyu Wang, Yi-Chen Zhang, Song Yu, Hong Guo
TL;DR
CV-QKD privacy amplification must process long inputs quickly while meeting finite-size security requirements across transmission distances. The paper develops a GPU-based length-compatible parallel implementation that divides long inputs into blocks and batches for FFT-accelerated processing. It reports an average speed of about 1.35Gbps at arbitrary input length, one to two orders of magnitude faster than previous implementations, with application to real-time CV-QKD.
Problem
High-speed privacy amplification is needed because CV-QKD postprocessing affects secret key rate and transmission distance, while finite-size security requires increasingly long inputs at longer distances.
Method
The paper implements length-compatible privacy amplification on a GPU by dividing long inputs into blocks and batches, using FFT acceleration and parallel batch processing.
Results
About 1.35Gbps average privacy-amplification speed is achieved at arbitrary input length, one to two orders of magnitude faster than the previous implementation.
Takeaways & Limitations
The proposed method supports high-speed real-time CV-QKD while accounting for finite-size effects in privacy amplification.
Takeaways & Limitations
The implementation assumes corrected keys are effectively identical after error verification, with authentication used to further ensure final-key identity.
Abstract
from arXiv · showhide
Privacy amplification is an indispensable step in postprocessing of continuous-variable quantum key distribution (CV-QKD), which is used to distill unconditional secure keys from identical corrected keys between two distant legal parties. The processing speed of privacy amplification has a significant effect on the secret key rate of CV-QKD system. We report the high-speed parallel implementation of length-compatible privacy amplification algorithm based on graphic processing unit. Length-compatible algorithm is used to satisfy the security requirement of privacy amplification at different transmission distances when considering finite-size effect. We achieve the speed of privacy amplification over 1 Gbps at arbitrary input length and the speed is one to two orders of magnitude faster than previous demonstrations, which supports high-speed real-time CV-QKD system and ensures the security of privacy amplification.
1. Introduction
CV-QKD uses postprocessing to generate secure keys, but privacy amplification speed is a bottleneck for high-speed real-time systems. This paper proposes a GPU-based length-compatible implementation to address finite-size requirements across transmission distances.
- Motivation: Privacy amplification distills final secret keys from identical corrected keys using universal hash families such as Toeplitz matrices.Toeplitz matrices have a simple structure that supports parallel implementation.
- Motivation: The postprocessing stage, especially high-speed privacy amplification, strongly affects CV-QKD secret key rate and transmission distance.
- Finite-size requirements: Finite-size security requirements make the required privacy-amplification input length increase with transmission distance, with a typical security parameter of 10^-10.
- Contribution: The paper proposes a high-speed GPU implementation of length-compatible privacy amplification for CV-QKD.The method extends earlier work applied in a high-speed real-time CV-QKD field test.
2. The Principle of Privacy Amplification for CV-QKD system
Privacy amplification uses universal hashing to compress corrected keys while limiting collision probability, with finite-size effects determining secure key length. The procedure applies a Toeplitz matrix to shared corrected keys, whose lengths and output rate are selected according to the security analysis.
- Universal hash families: Universal hashing extracts secure keys from corrected keys by selecting a hash function with low collision probability.Collision probability measures the chance that different inputs produce the same output key.
- Universal hash families: A Toeplitz matrix is a diagonal-constant universal hash function whose degrees of freedom are 2n − 1 rather than n^2.Its collision probability is n · 2^(-m+1), where n and m are the input and output lengths.
- Finite-size effect: The finite-size correction Δ(n) affects the final secret key rate and becomes more consequential as transmission distance increases.The analysis therefore favors large input blocks, while excessively long blocks become impractical to implement.
- Privacy-amplification procedure: The corrected-key strings uA and uB have length n, while the final keys rA and rB have length l = ⌊n × k⌋.Here, k is the secret key rate and the Toeplitz matrix has size n × l.
- Privacy-amplification procedure: Alice generates and sends a Toeplitz matrix, then both parties multiply their corrected keys by it to obtain final keys.The resulting keys can be treated as identical after error verification, with authentication further supporting their use.
- Length-compatible processing: The length-compatible algorithm divides shared weak keys and the Toeplitz matrix into blocks, independently hashes each block, and combines intermediate keys by modulo-2 addition.
3. High Speed Implementation of Length-Compatible Privacy Amplification
The method accelerates Toeplitz-matrix privacy amplification with FFT on a GPU while dividing long inputs into blocks and batches to accommodate finite-size requirements, GPU resources, and precision limits.
- Acceleration strategy: FFT reduces Toeplitz privacy-amplification complexity from O(n^2) to O(nlog^2n), making long-input processing more tractable.Direct processing becomes difficult for excessively long blocks, while FFT precision affects correctness and speed.
- Data partitioning: Long inputs are divided into blocks and smaller batches, with blocks processed serially and batches processed in parallel on the GPU.Intermediate keys are generated for each batch and combined by modulo-2 addition to obtain the final secret keys.
- Matrix partitioning: The Toeplitz matrix is partitioned into p row sub-matrices, while its columns remain undivided unless the final secret-key length requires further partitioning.The corresponding matrix representation is T = [T0, T1, · · · , Ti, · · · , Tp−1]′.
- Precision and compatibility: Length-compatible processing allows arbitrary input lengths while matching FFT precision to block size to avoid incorrect results.Single precision is sufficient below an FFT length of 2^23, whereas longer lengths require double or long-double precision.
- Parameter selection: Each block is further divided into q batches according to the input length and available GPU memory.The batch count and block count are selected from the privacy-amplification input size and GPU resources.
- Batch computation: Each batch is transformed with FFT, multiplied with its corresponding Toeplitz transform, processed with IFFT, and cropped to form an intermediate key.The selected IFFT outputs span positions n through n + k − 1.
4. Results
The GPU implementation uses length-compatible partitioning and single-precision computation to deliver high-speed privacy amplification across input lengths, while direct implementation is constrained by precision and GPU resources.
- Precision choice: Single precision is selected because half precision does not improve speed on the available GPU despite being supported.The implementation analyzes GPU resources, input length, and FFT precision as factors affecting speed.
- Measured performance: 1.35Gbps privacy-amplification speed is achieved at different input lengths on an NVIDIA TITAN Xp GPU.The secret-key length is 10% of the shared weak-key length in the reported table experiments.
- Direct implementation: Direct implementation reaches about 0.3Gbps between 4M and 80M input bits when double precision is used.Direct processing cannot be performed above 80M input bits because of GPU-resource limitations.
- Length-compatible implementation: Over 1Gbps speed is maintained at any input length with the length-compatible method, supporting finite-size security requirements and real-time CV-QKD.The blocks are processed serially and batches in parallel to handle inputs beyond direct GPU capacity.
- Comparison with prior work: The length-compatible implementation is faster than prior demonstrations reporting about 10Mbps, 108.77Mbps, and 65.443Mbps.These prior results used multiplication, number theoretic transform, and FPGA-based implementations, respectively.
5. Conclusions
The paper presents a GPU-based, length-compatible privacy amplification method for CV-QKD that supports arbitrary input lengths and accounts for finite-size security requirements. It achieves about 1.35 Gbps on average, reported as one to two orders of magnitude faster than previous implementations.
- The GPU-based method supports privacy amplification with arbitrary input length while considering finite-size effects for secret-key extraction.Long inputs are divided into blocks, accelerated with fast Fourier transforms, and processed through parallel batches.
- 1.35 Gbps average speed is achieved for arbitrary input length.
- One to two orders of magnitude faster than the previous implementation is reported.
- The early stage of the work was applied to the longest field test of a CV-QKD system.