Optimizing Robustness and Accuracy in Mixture of Experts: A Dual-Model Approach