Privacy vs Accuracy: Trade-offs in Federated Learning
Abstract
In federated learning, model updates from distributed sources are vulnerable to inference attacks that can reconstruct private training data from gradient information.Traditional approaches to this problem apply differential privacy uniformly across all model parameters, creating a significant trade-off where stronger privacy guarantees substantially degrade model utility.This thesis investigates how differential privacy can be selectively applied to gradient components to enhance privacy protection while maintaining model performance.We introduce selective gradient perturbation, a percentile-based approach that applies noise only to the most significant gradient components by magnitude.Through comprehensive experiments on the CIFAR-10 dataset using a simulated federated learning environment with 10 clients, we demonstrate that selective perturbation fundamentally transforms the conventional privacy-utility trade-off.Our findings reveal that by applying noise to only the top 10% of gradients, we achieve 74.38% accuracy while maintaining privacy protection, and counter-intuitively, stronger privacy settings (ε = 0.1) yielded better model performance (78.60% accuracy) than weaker protection (ε = 0.5, 66.56% accuracy).These results challenge the conventional wisdom that privacy necessarily comes at a significant cost to utility, offering a practical approach for implementing differential privacy in federated learning that achieves both strong privacy guarantees and high model…
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.