E2EGI: End-to-End Gradient Inversion in Federated Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Healthcare Informatics

Background:

• The proliferation of Internet of Medical Things (IoMT) generates vast healthcare data.
• Federated learning (FL) is crucial for preserving data privacy in distributed healthcare settings.
• Gradient Inversion Attacks (GIA) threaten FL by reconstructing data from leaked gradients, with prior limitations on batch size and label repetition.

Purpose of the Study:

• To propose a novel End-to-End Gradient Inversion (E2EGI) method to enhance Gradient Inversion Attacks (GIA).
• To improve the similarity of reconstructed samples and enable GIA on larger batch sizes and deep network models.
• To develop a new Label Reconstruction algorithm for improved accuracy using only gradient information.

Main Methods:

• Developed Minimum Loss Combinatorial Optimization (MLCO) for higher sample similarity.
• Implemented a Distributed Gradient Inversion algorithm for batch sizes from 8 to 256 on models like ResNet-50.
• Created a novel Label Reconstruction algorithm utilizing only gradient information.

Main Results:

• E2EGI achieved higher similarity in reconstructed samples compared to state-of-the-art methods.
• The Distributed Gradient Inversion algorithm successfully performed GIA on deep network models and large datasets (ImageNet).
• The Label Reconstruction algorithm attained 81% accuracy on a single batch with 96% label repetition, a 27% improvement.

Conclusions:

• The proposed E2EGI method significantly enhances the capabilities of Gradient Inversion Attacks in federated learning.
• This research provides a robust method for assessing data security risks in healthcare federated learning.
• The findings highlight the need for advanced privacy-preserving techniques in IoMT-driven healthcare systems.