What happens when every device in a federated network sees only one class of data?

Federated learning lets thousands of phones, sensors, or edge servers train a shared model without ever uploading their raw data. In the distillation variant, each device runs its local model on a public, unlabeled dataset and sends only the predictions back to a central server.

The trouble starts when local data is wildly uneven. One phone might have thousands of cat photos but zero dogs. Another sees only dogs. Each local model becomes biased toward whatever class dominates its own slice, and because the public dataset has no ground-truth labels, the server cannot easily tell which predictions are trustworthy. This label-distribution skew breaks the assumptions that most federated distillation methods rely on.

The 'unknown' class changes the game

FedCC adds a single extra output: an explicit 'unknown' tag. When a client's model is unsure about a public sample, it emits this tag instead of guessing among the real classes. The server then builds calibrated pseudo-labels on the public data using the remaining confident predictions, while the 'unknown' tags act as a pressure valve that prevents error propagation. No architectural changes are required on the client side. The local model still outputs a probability vector; the only difference is that the argmax can now land on the new index. The server aggregates these vectors exactly as before, then runs a calibration step that reweights the pseudo-labels based on how often each class appears across clients.

What the extreme skew experiment shows

A scenario where ten classes exist globally but each client holds samples from exactly one class represents the worst possible label distribution skew. Under these conditions FedCC reaches 67.3% top-1 accuracy on the global test set. Every baseline method tested collapses to near-random performance, which for ten classes means roughly 10%. The gap does not come from a more powerful model. It comes from refusing to hallucinate labels for classes the client has never seen. When a 'cat-only' client encounters a dog in the public set, its model naturally produces low confidence across all known classes. FedCC captures that low confidence as 'unknown' rather than forcing a cat label that would poison the server's pseudo-labels.

Calibration without ground truth

The server's calibration step works because the 'unknown' tags create a measurable signal: classes that receive many 'unknown' votes across clients are precisely the ones under-represented in local data. The algorithm boosts the pseudo-label weight for those classes and suppresses the weight for majority classes that dominate the confident predictions. This rebalancing happens entirely on the public dataset — no extra communication rounds, no access to private data. The authors note the method remains simple: one extra logit, one calibration pass per round. Code and hyperparameters are released alongside the paper. Globecom 2026 accepted the work for presentation.

Where the explanation stops

Wall-clock training time is not reported. Communication overhead per round is not reported. Memory footprint on constrained devices is not reported. Non-IID feature distributions — different lighting conditions, sensor noise, or adversarial corruptions — are also not tested. Whether the 'unknown' mechanism holds up in those settings is not addressed.