Federated open intent classification via granular-ball knowledge representation.

Open Intent Classification (OIC) aims to simultaneously identify known intents and detect unknown intents in open-world scenarios. However, existing OIC methods rely on centralized data collected from different devices, which is often infeasible due to privacy concerns. To address this, incorporate the federated learning paradigm into open intent classification, enabling both known intent recognition and unknown intent detection under decentralized settings. However, this setting still presents three core challenges: class inconsistency across clients, inadequate leverage of client-specific distributions, and uncontrolled generation of pseudo-unknown samples. To overcome these issues, we propose a Federated Open intent Classification method via Granular-Ball representations (FedOC-GB). During the training phase, each client constructs granular-ball representation to capture local class structures, generate structure-aware pseudo-unknown samples, and construct privacy-preserving local knowledge bases of known classes. On the server side, we aggregate model parameters and granular-ball knowledge from local clients to construct a global model and a unified knowledge base, enabling multi-granularity decision boundaries for open classification in inference. Comprehensive experiments on three benchmark datasets (CLINC, BANKING, and StackOverflow) demonstrate that FedOC-GB consistently outperforms state-of-the-art federated OIC methods, achieving up to 4.3% higher F1-All and 3.7% higher ACC on average. These results verify the effectiveness of our granular-ball representation in enabling privacy-aware and structure-preserving federated open intent recognition. The source code is publicly available at https://github.com/jiezhang64/FedOC-GB.