22 Nov 2026, Sun

Researchers Introduce Communication-Efficient Threshold-CKKS Protocol to Secure Cross-Silo Federated Learning

A team of cryptographic and machine learning researchers has introduced a novel protocol designed to overcome one of the most persistent bottlenecks in privacy-preserving collaborative artificial intelligence. In a newly published paper, authors Jihyung Kook, Dokyeong Kang, and Eunsang Lee present a communication-efficient threshold-CKKS protocol tailored specifically for cross-silo federated learning environments. The research addresses a fundamental tension in modern distributed machine learning: how to enable multiple organizations to train powerful models together using sensitive, private data without exposing the underlying information to servers or participating peers, and without incurring prohibitive network overhead.

Federated learning has rapidly evolved as a transformative paradigm for collaborative machine learning. By allowing multiple institutions—such as hospitals, financial banks, or corporate enterprises, often referred to as silos—to train shared predictive models locally without centralizing raw data, the technology respects data privacy regulations and organizational boundaries. However, privacy vulnerabilities remain a critical concern. Although raw data never leaves its origin, the model updates exchanged during the iterative training process can still leak sensitive information about the underlying datasets through sophisticated inference and reconstruction attacks.

To neutralize these threats, researchers have increasingly turned to homomorphic encryption, a cryptographic framework that allows computations to be performed directly on encrypted data without requiring decryption. While homomorphic encryption provides robust security guarantees, applying it to cross-silo federated learning has historically introduced severe practical hurdles. Specifically, existing systems struggle to simultaneously achieve collusion resistance, flexible model aggregation, graceful handling of client dropouts, and high communication efficiency. Many traditional schemes require a trusted key-distribution center, complicate server-side weight adjustments when participants drop offline during a training round, or demand massive volumes of data transmission that overwhelm network infrastructure.

The newly proposed protocol by Kook, Kang, and Lee seeks to eliminate these operational barriers through a sophisticated threshold-CKKS framework that operates entirely without a trusted key-distribution center. Under this architecture, independent local encryption keys are distributed among participating silos. This design choice empowers the central coordinating server to perform essential server-side weighting and manage dropout-aware reaggregation dynamically, all without forcing clients to retransmit ciphertexts when participants disconnect unexpectedly during a training cycle. By decoupling encryption keys from centralized control while preserving cryptographic security, the protocol offers a more resilient foundation for enterprise-scale distributed learning.

To address the severe communication overhead typically associated with homomorphic encryption, the researchers developed two distinct operational profiles known as Full and Sparse. The Full profile is engineered for dense model architectures and comprehensive updates, whereas the Sparse profile incorporates a Ring Learning With Errors to Module Learning With Errors conversion. This mathematical transformation enables fine-grained, chunk-wise aggregation while successfully retaining the compact packing properties characteristic of standard ring learning with errors schemes. Furthermore, the system leverages seeded ciphertext reconstruction, selective coefficient transmission, and compressed collective decryption techniques to systematically minimize the volume of data that must travel across the network during online phases.

The security foundations of the new protocol were rigorously examined under simulated conditions. The researchers established whole-view simulation security for fixed-cohort executions operating without dropouts or retries against a semi-honest server that might collude with up to a defined number of compromised clients. These security proofs hold under explicitly permitted leakage bounds and a strict minimum cohort requirement, ensuring that the threshold cryptographic mechanisms remain robust even in adversarial or untrusted server environments.

To evaluate the practical viability and performance of their approach, the authors conducted extensive empirical experiments using standard image classification benchmarks, specifically the CIFAR-10 and CIFAR-100 datasets. The experimental outcomes demonstrate a dramatic reduction in the performance gap traditionally observed between plaintext federated learning and encrypted alternatives.

According to the findings, the Full profile incurred only minimal accuracy penalties, registering just 0.32 and 0.29 percentage-point accuracy losses on CIFAR-10 and CIFAR-100, respectively. In terms of network utilization, the Full profile required only 1.78 times and 1.76 times the online communication volume of standard unencrypted plaintext training for the respective datasets.

The performance gains of the Sparse profile proved even more pronounced when evaluated against existing methodologies. When compared against FedPHE at matched accuracy levels, the Sparse profile drastically reduced online communication overhead by factors of approximately 79 and 84 for CIFAR-10 and CIFAR-100. Even when evaluated against accuracy targets set five percentage points below full-transmission plaintext models, the Sparse profile required only modest network resources—specifically 2.5 times and 2.1 times the online communication of unencrypted plaintext baselines.

By demonstrating that encrypted federated learning can achieve high classification accuracy while bringing network communication costs remarkably close to plaintext standards, the research marks a significant step forward. The work provides a scalable, secure, and communication-efficient pathway for organizations seeking to harness collaborative artificial intelligence without compromising data confidentiality or operational efficiency.

Leave a Reply

Your email address will not be published. Required fields are marked *