New PI-SME method enhances gradient inversion attacks in federated learning

Researchers have developed a new method called the Path-Integral Surrogate Model Extension (PI-SME) to improve gradient inversion attacks in federated learning. This technique treats a client's model update as a path integral of the gradient field, approximating it using Gauss--Legendre quadrature along a learnable Bézier curve. PI-SME aims to reconstruct private input data more accurately than existing methods by analyzing multiple points along the trajectory between the client's initial and final model states. Experiments on CIFAR-100 and FEMNIST datasets demonstrated PI-SME's effectiveness in faithfully reconstructing private inputs across various conditions. AI

IMPACT This research highlights potential privacy vulnerabilities in federated learning, prompting further investigation into robust defense mechanisms.

RANK_REASON The cluster contains a research paper detailing a new method for gradient inversion attacks in federated learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Bézier curve
• CIFAR-100
• FedAvg
• federated learning
• FEMNIST
• Path-Integral Surrogate Model Extension
• PI-SME

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →