Morning Edition · Thursday, June 18, 2026Published at 6:48 AM EDT · New York

TIGER Recovers Training Inputs From Federated Learning Gradients

A new attack reconstructs private inputs by optimizing distances in embedding subspaces, sharpening doubts about gradient sharing.

A paper posted to arXiv presents TIGER, a gradient inversion attack that reconstructs private inputs from the gradient updates clients send in federated learning. The method works by inverting transformer gradients through optimization over…