A privacy-preserving vertical federated learning method resistant to collusion and tolerant to straggling nodes

By combining Epoch-level batch mask generation and Diffie-Hellman key sharing with Shamir secret sharing and Lagrange interpolation recovery, the collusion attack and lagging problem in vertical federated learning is solved, achieving strong resistance to collusion, lagging tolerance and efficient and secure aggregation, thus improving the system's robustness and training efficiency.

CN122389079APending Publication Date: 2026-07-14UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
UNIV OF ELECTRONICS SCI & TECH OF CHINA
Filing Date
2026-05-15
Publication Date
2026-07-14

AI Technical Summary

Technical Problem

In practical deployments, vertical federated learning struggles to simultaneously defend against malicious collusion attacks by multiple participants and the problem of laggards, making it difficult to coordinate the optimization of privacy security and system robustness. It also results in high computational and communication overhead and compromised model accuracy.

Method used

Employing a combination of techniques including Epoch-level batch mask generation, Diffie-Hellman pairwise session key sharing, Shamir secret sharing, and Lagrange interpolation recovery, this approach achieves strong anti-collusion capabilities, fallback tolerance, and efficient verifiable secure aggregation through zero-sum mask pre-generation, secure client-side embedded computation, secret sharing homomorphic aggregation, and server-side embedded recovery.

Benefits of technology

It achieves strong privacy protection against multi-client collusion under arbitrary thresholds, tolerates laggards, reduces communication and computational overhead, maintains model accuracy, and improves system robustness and training efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122389079A_ABST
    Figure CN122389079A_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of information security, privacy computing and distributed machine learning, and discloses a privacy protection longitudinal federated learning method resisting conspiracy and tolerating lag nodes. The method aims to solve the technical problem that the privacy security and the recovery threshold are strongly bound in the traditional secret sharing scheme, so that the privacy protection and the robustness are difficult to be optimized simultaneously. The method comprises the following steps: each client generates a session key through a key exchange protocol, and generates a zero-sum mask matrix based on a batch index; the original embedding of the local feature is forward propagated, the mask embedding is obtained by adding the mask matrix to the quantization mapping, and the share distribution is constructed by constructing a polynomial; after receiving all the shares, the aggregated shares are uploaded to the server by using the homomorphism characteristic of addition; the server collects the qualified shares, reconstructs the global embedding sum by interpolation, offsets the mask by using the zero-sum characteristic, and obtains the global average embedding by dequantization, and inputs the top model to calculate the loss and gradient to update the client bottom model.
Need to check novelty before this filing date? Find Prior Art