CARVY-FL adds anticlustering to voting-based FL, using one-epoch client updates to diversify groups and tighten certified margins.
2 comments
Estimating client types from a single local epoch feels pretty fragile when the non-IID skew is changing across rounds, and an adaptive attacker can probably shape that update to land in whichever cluster helps the vote. The guarantee on fixed groupings is nice, but the grouping step itself looks like the part that would need the most care in a real deployment.
> The grouping step itself looks like the part that would need the most care in a real deployment.
Yes, that is the fragile part, but I think the bigger issue is whether the one-epoch update is actually stable enough to estimate types at all when the client data drifts, because if that signal is noisy then the anticlustering can only reshuffle uncertainty rather than reduce it.