Balancing Security, Accuracy, and Adaptability for CNNs in Federated Learning
Ansh Chauhan, Deepti Bhat, S Chirashwi, Chiranth Nagaprasanna, Swathi B H, D N Sachin · 2025
Federated Learning (FL) allows decentralized training of models while maintaining privacy by keeping data on local devices. However, achieving a balance between privacy, adaptability, and accuracy remains challenging in non-IID settings. This paper presents a novel algorithmic stack that integrates Differential Privacy (DP-SGD), Shuffling, FedMedian aggregation, and the Reptile meta-learning algorithm to train Convolutional Neural Networks (CNNs) securely and effectively across clients. Experimental results on MNIST, Fashion-MNIST, and CIFAR-10 datasets demonstrate significant improvements: MNIST achieved 98.15% accuracy for 10 clients, Fashion-MNIST 84.80%, and CIFAR-10 73.54%. The proposed stack ensures enhanced robustness against adversarial attacks, improved generalization across heterogeneous data, and rigorous privacy guarantees. This framework addresses key limitations in existing FL research by uniting privacy, security, and adaptability in a single robust solution.