Tuesday 11 March 2025
A novel approach to federated learning, a technique used to train AI models on decentralized data, has been proposed by researchers. The method, which involves clustering local model parameters and updating them based on global centroids, has been shown to converge to the Federated Averaging (FedAvg) global model as the number of clusters increases.
Federated learning is a powerful tool for training AI models on large datasets that are distributed across multiple devices or organizations. However, it can be challenging to implement, particularly when dealing with sensitive data that needs to remain private. The traditional approach to federated learning involves aggregating the local model parameters from each device or organization and then updating them using a centralized algorithm.
The new method proposed by researchers takes a different approach. Instead of aggregating the local model parameters directly, they are first clustered into groups based on their similarity. This is done using k-means clustering, a popular unsupervised machine learning algorithm. The centroids of these clusters are then used to update the local model parameters.
The researchers have shown that this approach can converge to the FedAvg global model as the number of clusters increases. This means that the updated local model parameters will become increasingly similar to the global model, even though they were only trained on a small portion of the overall dataset.
One of the key benefits of this approach is that it reduces the amount of data that needs to be transmitted between devices or organizations. In traditional federated learning, all of the local model parameters need to be transmitted to the central server in order to be updated. With the new method, only the centroids of the clusters need to be transmitted, which can significantly reduce the amount of data that needs to be transferred.
Another advantage of this approach is that it provides an additional layer of security for sensitive data. Because the local model parameters are clustered and updated based on global centroids, there is less risk of sensitive information being leaked or compromised during the training process.
The researchers have tested their method using a range of datasets and models, including image classification tasks and natural language processing applications. The results show that the new approach can achieve similar levels of accuracy to traditional federated learning methods, while reducing the amount of data that needs to be transmitted and providing additional security benefits.
Overall, this novel approach to federated learning has the potential to revolutionize the way AI models are trained on decentralized data.
Cite this article: “Clustering-Based Federated Learning Method for Decentralized Data Training”, The Science Archive, 2025.
Federated Learning, K-Means Clustering, Unsupervised Machine Learning, Ai Models, Decentralized Data, Local Model Parameters, Global Centroids, Data Transmission, Sensitive Information, Security Benefits.
Reference: Ali Abbasi Tadi, Dima Alhadidi, Luis Rueda, “Trustformer: A Trusted Federated Transformer” (2025).
