<p>This paper proposes a multi-objective joint optimization framework (MOJO) for text classification, aiming to synergistically enhance model performance, fairness, and privacy protection. Existing methods predominantly focus on single objectives, struggling to achieve balanced multi-objective outcomes in complex application scenarios. MOJO employs a dynamic dual-path fusion mechanism, integrating manifold regularization with structure-preserving adversarial strategies to capture complementary global and local features. To balance privacy and semantic fidelity, it introduces reconstruction error-driven weighted fusion and attention-guided perturbation control, alongside a noise injection strategy to enhance privacy protection. Concurrently, label-guided structural perturbations and trajectory contrastive learning improve fairness and robustness in the representation space. Experiments demonstrate that MOJO significantly outperforms existing methods across multiple benchmark tasks, achieving synergistic optimization of classification performance, fairness, and privacy protection.</p>