The model is wrong. The data is private. Debug without looking.
Federated learning: multiple parties train a shared model without sharing their data. When the model underperforms, debugging requires identifying which data contributed to the failure. But inspecting the data violates the privacy guarantee that justified federation in the first place (arXiv:2603.21650).
The federated debugging framework diagnoses data quality issues — label noise, distribution shift, feature corruption — without centralized data access. The approach uses model behavior as a proxy for data inspection: how the global model's predictions change when different parties' contributions are included or excluded reveals which parties' data is problematic. Gradient-based influence functions estimate each party's contribution to specific errors without transmitting the data itself.
The diagnosis is necessarily approximate — you can tell which party has problematic data but not which specific examples are wrong. The privacy guarantee constrains the debugging resolution. You get party-level attribution, not example-level attribution.
The structural insight: debugging under privacy constraints is a fundamentally different problem from debugging with full access. With full access, you inspect the data, find the corrupt examples, fix or remove them. Under privacy, you can't inspect — you infer. The inference is indirect: model behavior → gradient influence → party attribution → suspected data issue. Each step adds uncertainty. The debugging resolution is set by the privacy constraint, not by the debugging technique. More privacy means coarser debugging. The tradeoff is irreducible — seeing less means diagnosing less precisely.