Privacy-Preserving Image Classification based on Federated Learning with Hybrid CNNs Model on MNIST Data

Background: Machine learning (ML) privacy problems have prompted the creation of privacy-preserving methods, one of which is Federated Learning (FL), which has emerged as an important solution. FL protects data confidentiality by enabling decentralized client devices to train models collaboratively without sharing raw data. Methods: This study uses a Hybrid Convolutional Neural Network (CNN) that combines convolutional layers for feature extraction with fully connected layers for classification to apply FL for image classification using the MNIST dataset. A number of neural networks are assessed, such as a Hybrid CNN, Multilayer Perceptron (MLP), and Recurrent Neural Network (RNN). Results: The Hybrid CNN achieved a maximum accuracy of 93.78% and a minimum loss of 0.0034, outperforming MLP (90.37% accuracy) and RNN (53.68% accuracy). Conclusion: These results demonstrate that Hybrid CNNs are more effective for image classification in FL settings than conventional models, making them an optimal choice for privacy-preserving ML applications.