Paper 2026/2897
Communication-Efficient Threshold-CKKS Aggregation for Secure Cross-Silo Federated Learning
Abstract
Federated learning (FL) enables collaborative training without sharing raw data, but exchanged model updates may reveal sensitive information. Homomorphic encryption (HE) protects these updates, yet combining collusion resistance, flexible aggregation, dropout handling, and communication efficiency remains challenging. We propose a communication-efficient threshold-CKKS protocol for cross-silo FL with Full and Sparse profiles, without a trusted key-distribution center. Independent local encryption keys enable server-side weighting and dropout-aware reaggregation without ciphertext retransmission. Sparse uses RLWE-MLWE conversion for fine-grained chunk-wise aggregation while retaining compact RLWE packing. Seeded ciphertext reconstruction, selective coefficient transmission, and compressed collective decryption further reduce communication. We establish whole-view simulation security for fixed-cohort executions without dropout or retry against a semi-honest server colluding with up to $c$ clients, under explicitly permitted leakage and a minimum cohort requirement of $\tau\ge c+2$. Experiments on CIFAR-10 and CIFAR-100 show that Full incurs only 0.32 and 0.29 percentage-point accuracy losses, with 1.78$\times$ and 1.76$\times$ the online communication of plaintext, respectively. At matched accuracy levels, Sparse reduces online communication by factors of approximately 79 and 84 compared with FedPHE. At accuracy targets five percentage points below full-transmission plaintext, Sparse requires only 2.5$\times$ and 2.1$\times$ the online communication of plaintext, respectively. These results narrow the communication gap between encrypted and plaintext FL.
Metadata
- Available format(s)
-
PDF
- Category
- Cryptographic protocols
- Publication info
- Preprint.
- Keywords
- Federated learningHomomorphic encryptionThreshold CKKSSecure aggregationCollective decryption
- Contact author(s)
-
jhkook30 @ gmail com
s123kdg @ gmail com
eslee3209 @ sejong ac kr - History
- 2026-11-19: approved
- 2026-11-15: received
- See all versions
- Short URL
- https://ia.cr/2026/2897
- License
-
CC BY-NC-ND
BibTeX
@misc{cryptoeprint:2026/2897,
author = {Jihyung Kook and Dokyeong Kang and Eunsang Lee},
title = {Communication-Efficient Threshold-{CKKS} Aggregation for Secure Cross-Silo Federated Learning},
howpublished = {Cryptology {ePrint} Archive, Paper 2026/2897},
year = {2026},
url = {https://eprint.iacr.org/2026/2897}
}