Quantifying and Defending against the Privacy Risk in Logit-based Federated Learning

Name: Quantifying and Defending against the Privacy Risk in Logit-based Federated Learning
Uploaded: 2026-10-22T22:56:31.611078+00:00
Duration: 647 s
Description: This research analyzes a hidden privacy risk in logit-based Federated Learning where a semi-honest server can learn private models from training logits. It introduces AdaMSA to quantify this risk by simulating an adaptive model stealing attack and proposes FedLP, a perturbation defense strategy, to protect client models while maintaining utility.

This research analyzes a hidden privacy risk in logit-based Federated Learning where a semi-honest server can learn private models from training logits. It introduces AdaMSA to quantify this risk by simulating an adaptive model stealing attack and proposes FedLP, a perturbation defense strategy, to protect client models while maintaining utility.

Open the full episode