Paper screens federated updates against label-flipping attacks
Original title: Peer-Relative Trust-Guided Update Screening for Robust Semi-Decentralized Federated Learning under Label-Flipping Attacks
On 50 MNIST clients, screening rejected most malicious updates while keeping most benign ones.
A paper extends a trust-based semi-decentralized federated learning framework with peer-relative update screening that combines a validation-utility score with direction- and norm-based quality checks. The evaluation uses 50 MNIST clients, 20 percent persistent label-flip attackers, Dirichlet heterogeneity of 0.5 and 0.1, 100 rounds, and five paired seeds. Under moderate heterogeneity it reports 98.37 percent accuracy, rejecting 95.82 percent of malicious updates and retaining 90.30 percent of benign ones. Under severe heterogeneity accuracy is 95.25 percent, with lower rejection and retention; trust weighting alone accepted every malicious submission.
Unconfirmed: Reported rates are the authors' MNIST simulation results, not observations from a production system.
- Moderate-heterogeneity trust distributions were fully separated, with KS of 1.00.
- Severe heterogeneity retained 76.90 percent of benign updates and KS of 0.96.
- Authors position the method as an off-chain governance core for later blockchain layers.
From the original article
1 min read at arxiv.orgTrust and reputation mechanisms can reduce the influence of malicious clients in federated learning (FL), but strict update filtering is difficult under non-independent and identically distributed (non-IID) data because honest local updates may temporarily reduce performance on a shared validation proxy.
By Ajay Kumar Shrestha · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.