FGF-CNN: Federated Gradient Fusion For Convolutional Neural Networks – A Collaborative Learning Approach

Federated learning is a type of technique that makes it possible to train the models using a centralized process, at the same time ensuring that individual privacy is observed. We also propose the Federated Gradient Fusion Convolutional Neural Network (FGF-CNN) to which the gradient fusion technique is applied for the global integration of local architectures in this paper. The FGFCNN is intended for image classification tasks through the use of a CNN inclusive of two dimensional convolution layer, the Max Pooling layer, flattening process as well as a fully connected softmax layer. The proposed model was trained under the SugarCane dataset which harbors several types of sugarcane image classes. The method being proposed uses client side gradient updates that make the results more generalizable than the raw data used which in turn helps with privacy considerations. The results lead to high accuracy and efficiency in federated environments with a lower number of exchanged messages counted as a communication overhead, which makes FGF-CNN applicable to real-world distributed learning setups where the data privacy is a concern.