Federated Learning Privacy Protection via Training Randomness
Abstract
Federated learning is a collaborative machine learning paradigm that trains models across multiple computing nodes while aiming to preserve the privacy of local data held by participants. However, because of the open network environment, federated learning faces severe privacy and security challenges. Studies have shown that attackers can reconstruct original training data by intercepting gradients transmitted across the network, thereby posing a serious threat to user privacy. One representative attack is the Deep Leakage from Gradients (DLG), which iteratively recovers training data by optimizing dummy inputs to match the observed gradients. To address this challenge, this paper proposes a novel privacy-preserving strategy that leverages the randomness inherent in model training. Specifically, during the training process, we introduce the dropout method to selectively disconnect neuronal connections, while adding Gaussian noise to mask the gradients of the disconnected neurons. This approach effectively disrupts gradient-leakage attacks and reduces overfitting, thereby enhancing model generalization. We further provide a theoretical analysis of privacy guarantees using differential privacy metrics. Extensive experiments under federated-learning attack-defense scenarios demonstrate the effectiveness of the proposed strategy. Compared with existing defenses, our method achieves strong privacy protection against gradient leakage while maintaining competitive model accuracy,…
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.