Federated Learning Client Parameter Verification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning systems face poor robustness due to Byzantine attacks, where malicious data from compromised clients can propagate incorrect model parameters, affecting all clients and leading to poor training results.
Innovation Solution
Implement a method where clients verify model parameters received from the server against local values, using training data and performance metrics to determine if the parameter meets certain conditions, and if not, they use local values for training, sending alarm information to the server to prevent propagation of incorrect data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the server delivers model parameters to all clients in a federated learning system, then the training process can proceed, but Byzantine attacks can propagate malicious parameters to all clients causing poor robustness
Solution Approach 1:
The patent implements preliminary verification of model parameters before they are used for training. Each client checks whether the received parameter satisfies a preset condition (e.g., within expected range or consistent with local model state) before applying it. This preliminary action prevents malicious parameters from corrupting the training process, thereby improving system robustness without requiring changes to the overall federated learning architecture.
2Reliability
If the server performs detection of malicious data, then security can be improved, but the detection load and computational overhead increase
Solution Approach 1:
The patent implements a self-service mechanism where each client independently verifies the validity of received model parameters using locally available information and preset conditions. This distributes the detection burden from the server to the clients themselves, maintaining security while significantly reducing the server's computational load and energy consumption. The client autonomously determines whether to use the received parameter or fall back to local training.
3Productivity
If clients use received parameter values for training, then training efficiency is maintained, but accuracy of training results deteriorates when parameters are malicious
Solution Approach 1:
The patent applies preliminary verification of parameter validity before training execution. Each client checks whether the received parameter meets preset conditions (such as being within a reasonable range or consistent with the local model state). Only parameters that pass this preliminary check are used for training, ensuring both efficiency (by using received parameters when valid) and accuracy (by rejecting malicious parameters).
Solution Approach 2:
The patent implements a feedback mechanism where clients monitor the effectiveness of training using received parameters and report anomalies to the server. If a client detects that a received parameter leads to poor training performance or violates expected constraints, it can flag this feedback to the server, which can then adjust the parameter distribution strategy or investigate potential Byzantine attacks, thereby improving overall training accuracy while maintaining efficiency.
Data Source
AI summary
A federated learning method, device, and system are provided, to improve robustness of the federated learning system. The method includes: A first client receives a first value of a parameter of a machine learning model from a server, where the first client is one of a plurality of clients; when the first value of the parameter does not meet a first condition, the first client performs a current round of training based on first training data, the machine learning model, and a local value of the parameter, to obtain a training result of the current round of training, where the first training data is data reserved on the first client; and the first client sends the training result and alarm information to the server, where the alarm information indicates that the first value of the parameter does not meet a requirement.