Federated learning trains models across distributed clients without centralizing data, but malicious participants can undermine the process by flipping labels on their local data. Trust and reputation mechanisms are often used to reduce the influence of such attackers. However, as the paper notes, strict update filtering becomes difficult under non-IID data, because honest clients' updates can legitimately diverge from the norm.
The authors propose a peer-relative trust-guided update screening method for semi-decentralized federated learning. Rather than applying a fixed global threshold, the approach evaluates each client's update relative to its peers, aiming to filter out malicious contributions while preserving honest ones that merely reflect local data distribution.
The abstract is truncated, so full experimental results are not available in the source. Still, the motivation is clear: robust defense against label-flipping requires separating true malicious behavior from the natural heterogeneity of honest clients, and peer-relative trust is a promising way to do that.