Federated Learning Privacy Gradient Buffering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In federated learning for recommendation models, ensuring user privacy and security is challenging, especially when user terminal devices may drop out, as existing methods fail to effectively protect data privacy and maintain model performance in distributed training scenarios.
Innovation Solution
A method for distributed privacy-preserving learning that involves each user terminal device locally acquiring a scoring matrix, iteratively training item and user embedding matrices, and using noise addition to ensure differential privacy, with a server determining surviving devices and updating matrices accordingly to maintain model accuracy and privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If federated learning is used for distributed training, then data privacy is protected, but model training reliability deteriorates when user terminal devices drop out
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing gradient information in a gradient buffer before device dropouts occur. The system accumulates gradients from multiple iterations and stores them in the buffer, so that when a device drops out, the pre-stored gradient information can be used to continue training without restarting, thus maintaining training reliability while preserving privacy through federated learning
Solution Approach 2:
The patent implements beforehand cushioning by introducing a gradient buffer that acts as a cushion against device dropouts. The buffer stores gradient information in advance, providing a cushioning effect that absorbs the impact of sudden device departures and allows the training process to continue smoothly without compromising either privacy protection or training reliability
2Reliability
If noise is added to protect differential privacy, then data security is improved, but model training precision deteriorates
Solution Approach 1:
The patent applies partial action by selectively adding noise only to the gradient information that is stored in the gradient buffer, rather than adding noise to all training data and parameters. The noise is added with a carefully controlled variance that provides sufficient privacy protection while minimizing the impact on model training precision, thus achieving a balance between security and precision
3Adaptability or versatility
If user terminal devices are allowed to drop out, then system adaptability is improved, but data privacy protection deteriorates
Solution Approach 1:
The patent introduces the gradient buffer as an intermediary between the training process and device dropouts. The buffer mediates the interaction by storing gradient information in a privacy-preserving manner, allowing devices to join and leave freely while the buffer maintains the integrity and privacy of the accumulated gradient information, thus enabling adaptability without compromising privacy protection
Data Source
AI summary
A target user terminal device among a plurality of user terminal devices performing federated learning acquires a scoring matrix corresponding to a target user, and performs the following operations in at least one iteration of training an item embedding matrix and a user embedding matrix corresponding thereto: determining a loss function gradient corresponding to the item embedding matrix in a current iteration, and sending to a server the loss function gradient added with a first noise; if a survival notification sent by the server is received, sending to the server a second noise for reducing the noise added to the loss function gradient, so that the server updates the item embedding matrix based on an aggregation result of the loss function gradient added with the first noise and the second noise, and sends an update result to each of the user terminal devices.