Synchronous federated learning is only as fast as its slowest client, and stragglers arise both from weak local compute and from an unreliable wireless link. This work attacks both causes at once: a digital-twin network at the server mirrors each client's insensitive data so part of the local training can be executed server-side, and non-orthogonal multiple access lets every selected client upload its parameters over the same channel simultaneously, with successive interference cancellation at the receiver. Unlike earlier digital-twin FL frameworks that assume the twin can mirror a client's whole dataset, here the mapped fraction is capped by the client's insensitive-data share, so privacy is preserved and the aggregation weight splits between real local data and mapped data. On top of that, a reputation-based client selection scheme scores each client on three axes - accuracy contribution modelled as a concave Weibull function of data size, model staleness in the age-of-update sense, and the ratio of positive to negative interactions as judged by a reject-on-negative-influence detector - so that poisoning clients are screened out before aggregation without adding selection complexity. The resource-allocation problem is then posed not as a weighted sum but as a Stackelberg game that makes the competing objectives explicit: the clients lead, minimising total energy over transmit power, CPU frequency and mapping ratio, while the server follows, minimising total latency over the server frequency coefficients it grants each client. The follower's optimum is shown to equalise all clients' digital-twin computing times, and substituting that best response reduces the leader problem to subproblems solved in closed form, with a Dinkelbach algorithm handling the concave-convex fractional power subproblem and KKT conditions closing it out. Simulations on IID and non-IID MNIST and CIFAR-10 vary the number of poisoners and the twin's mapping deviation.
Despite the advantage of preserving data privacy, federated learning (FL) still suffers from the straggler issue due to the limited computing resources of distributed clients and the unreliable wirel…