Fed-RoD (Federated Robust Decoupling for Generic and Personalized Learning)

A dual-objective federated classification framework (Chen & Chao 2022, Section 4, Equations 6-9, Theorem 1) that simultaneously trains a class-balanced generic model and personalized client predictors without extra communication rounds by decoupling the network into a shared feature extractor f(x; theta), a shared generic linear head h^G(z; W_G) trained with Balanced Softmax Risk Management (BSRM, Eq. 7: l_BSRM(x, y) = -log(N_{m,y}^gamma exp(g_y) / sum_{c'} N_{m,c'}^gamma exp(g_{c'}))), and a lightweight personalized head h^P(z; v_m) whose logits are added to the generic logits via hat_y^P = g(x; theta, W_G) + h^P(f(x; theta); v_m) and trained with empirical cross-entropy on client m's local class prior.