Latent Entity Negotiation for Federated Learning with Private Taxonomies

Federated classifiers normally assume that output coordinate $j$ denotes the same class at every client. This interface breaks when institutions retain taxonomy-local labels that independently permute, merge, or conditionally split latent entities; here “private” means unshared identities, not differential privacy. We introduce FedLENS, an architecture that factors each private classifier into a shared latent-entity bank, a client-private entity-to-word channel, and an input-dependent residual reader. Clients communicate adaptive micro-prototypes without word identities, and the server negotiates anonymous latent entities through mass-aware semi-unbalanced transport with learned capacities. The residual preserves feature-conditioned refinements, while the entity path transfers structure through the encoder even though the default test-time reader is private. We prove vocabulary-permutation invariance, conditional local identifiability under separated-support, pure-component, and correct-capacity assumptions, and conditional encoder convergence under stochastic-gradient local updates with explicit private-state and bank-value variation terms. Across four controlled benchmarks, negotiation improves five-seed mean accuracy by 5.00–7.62 points over the matched residual-only control; gains over the strong native-compatible FedRoD comparator are smaller, at 0.23–1.05 points.