Federated learning (FL) enables collaborative training across distributed clients without centralizing raw data, making it an attractive approach for privacy-sensitive applications. However, shared model updates in FL may still leak information, leaving systems vulnerable to inference attacks. Differential privacy (DP) provides formal guarantees but often degrades performance, especially in non-independent and identically distributed (non-IID) settings. This work proposes an adaptive noise scaling mechanism to integrate DP into FL more effectively. The method dynamically adjusts client-level noise based on local loss variance, balancing privacy preservation and model utility across heterogeneous clients. In addition, an enhanced Convolutional Neural Network (CNN) architecture with Group Normalization and residual connections is employed to stabilize training and improve generalization under noisy updates. Experiments on the MNIST dataset with 50 clients show that the adaptive federated DP model achieves 96.16% accuracy with a privacy budget of at a noise multiplier of 1.0. This performance surpasses the centralized DP baseline (94.15%) while approaching the non-private FL baseline (99.57%). Overall, the results highlight adaptive differential privacy as a practical and scalable approach for privacy-preserving federated learning, with strong potential in domains such as healthcare, finance, and mobile edge computing.
Keywords: adaptive noise scaling, convolutional neural networks, differential privacy, federated learning, privacy-utility trade-off
