A decentralized federated learning secure aggregation method and system
By using multi-view screening and reinforcement learning algorithms to evaluate neighbor node models, the Byzantine attack problem in decentralized federated learning is solved, improving the model's security and performance, and ensuring its robustness and accuracy.
CN122132725APending Publication Date: 2026-06-02HUNAN UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUNAN UNIV OF SCI & TECH
- Filing Date
- 2026-02-02
- Publication Date
- 2026-06-02
Smart Images
Figure CN122132725A_ABST
Abstract
This invention discloses a decentralized federated learning secure aggregation method and system. The method specifically includes: each worker node training on its local training dataset to obtain a local model; each worker node interacting with neighboring nodes based on the communication topology to obtain the neighboring nodes' local models; each worker node evaluating the received neighboring node local models from multiple dimensions based on its own local model, and filtering the neighboring node local models using a reinforcement learning algorithm; each worker node aggregating the filtered neighboring node local models with its own local model to update its local model; and iteratively updating the training until a preset maximum number of communication rounds is met. This invention filters toxic models through multi-perspective screening and optimizes evaluation metrics to select the local models to be aggregated, thereby increasing the difficulty for attackers and trapping them in an adversarial dilemma.
Need to check novelty before this filing date? Find Prior Art