Trustworthy and privacy-preserving decentralized federated learning with multi-layer defense for secure collaborative AI

Short summary

A new decentralized federated learning (DFL) framework integrates hybrid privacy (public pretraining + differential privacy) and multi-layer defenses (data sanitization, peer verification, robust aggregation, inference protection) to counter both privacy leakage and poisoning/backdoor attacks, improving clean accuracy from 83.10% to 86.12% and adversarial accuracy under FGSM, PGD, and CW attacks by up to 30 percentage points.

AI-generated from the title and abstract; the full text is not read.