Federated Unlearning via Malicious Client Update Subtraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing federated unlearning methods are inefficient in handling malicious clients, leading to delayed training and inaccurate model predictions due to the inability to effectively eliminate the influence of inferior data provided by such clients, and they often require retraining, which is time-consuming and inefficient.
Innovation Solution
A federated unlearning method that subtracts parameter updates of malicious clients from the global model, using a central server to perform unlearning operations without client cooperation, and includes a comparison mechanism to terminate unlearning operations prematurely to prevent excessive degradation, utilizing a benchmark dataset for performance recovery.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional federated unlearning method (retraining) is used to exclude malicious clients, then the model can eliminate malicious influence, but the training time is greatly extended and efficiency is reduced
Solution Approach 1:
The patent records parameter updates of each client during the original FL training process. When unlearning is needed, the server directly subtracts the malicious client's recorded parameter updates from the global model, eliminating the need to retrain from scratch. This preliminary recording action enables fast unlearning while maintaining model accuracy.
Solution Approach 2:
The patent extracts and isolates the parameter updates contributed by malicious clients during the original training. By identifying and separating these specific parameter updates, the system can remove only the malicious influence without retraining the entire model, thus reducing time loss while maintaining reliability.
2Reliability
If federated unlearning is performed during the training round when client requests erasure, then the client's contributions are erased, but subsequent FL training is delayed
Solution Approach 1:
The system performs unlearning operations using pre-recorded parameter updates independently of the FL training schedule. This allows unlearning to occur without interrupting or delaying subsequent FL training rounds, as the unlearning computation is based on previously stored data rather than requiring active client participation during training.
Solution Approach 2:
The patent uses recorded parameter updates as an intermediary representation of client contributions. Instead of requiring direct client participation for unlearning, the system mediates the process by subtracting these recorded updates from the global model, enabling unlearning to proceed independently and without delaying FL productivity.
3Ease of operation
If existing federated unlearning methods are used, then unlearning operation can be performed, but they cannot effectively eliminate malicious influence when client provides inferior data
Solution Approach 1:
The patent introduces a verification mechanism where the server evaluates the quality of parameter updates received from clients during the original FL training. By monitoring changes in global model performance and comparing parameter update patterns, the system can identify malicious clients whose inferior data degrades model performance, and subsequently apply unlearning operations targeted at these identified malicious contributions.
Solution Approach 2:
The system extracts and identifies parameter updates from malicious clients by analyzing the impact of each client's contributions on overall model performance. Once malicious parameter updates are identified and separated, the unlearning operation can be precisely applied to remove only the harmful influence while preserving legitimate client contributions, thereby improving both the ease of operation and effectiveness of unlearning.
Data Source
AI summary
A federated unlearning method based on malicious terminal intervention training and belongs to the technical field of privacy computing and federated learning, which eliminates the influence of the malicious client on the global model through federated unlearning and subtracts the parameter updates of the malicious client from the parameters of the final global model generated by federated learning to save the retraining time by continuing training with a theoretically unusable low-quality model. A comparison mechanism for judging the effect of the previous round of unlearning model and the effect of the current round of unlearning model to analyze the unlearning effects is also provided. The final unlearning model is trained with a small dataset and the deviations produced by the training process on the model are recovered, which effectively improves the accuracy of the final unlearning model.