Defending Patient Privacy in Federated Learning with Shadow Model
Abstract
Abstract Federated Learning (FL) has become a approach for training models together in privacy-sensitive domains like healthcare, where sharing of raw data is frequently restricted. But recent research has shown that Gradient Inversion Attacks (GIAs) can use shared model updates to recreate sensitive training data, which puts patient privacy at risk. Current defense mechanisms, like differential privacy and gradient perturbation techniques, often use uniform protection strategies that either make the model work less well or don't protect important data areas well enough.This study presents an enhanced privacy-preserving framework based on an augmented Shadow Defense (Shadow Def) mechanism to address these limitations.The suggested method includes a multi-phase, region-aware defense strategy that uses Fast Fourier Transform for frequency-domain sensitivity analysis, Mean Squared Error (MSE) for vulnerability mapping, and multi-component noise injection. To make the system more resistant to reconstruction attacks while keeping training stable, gradient direction perturbation, temporal smoothing, and adaptive noise scaling are also used. The suggested method has undergone comprehensive testing in a federated learning context, utilizing simulated gradient inversion attacks on two prominent medical imaging datasets: Chest X-Ray and Eye PACS. The adversarial training process significantly reduces reconstruction quality (i.e., reconstruction errors), as shown by a higher RMS error…
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.