摘要： ：针对多语言机器翻译中数据隐私保护问题，研究提出了一种结合差分隐私机制的多语言文本到文本转换模型（mT5）翻译模型，旨在保护用户隐私的同时维持翻译质量。首先，在模型微调阶段引入梯度裁剪操作，对每个样本的梯度施加范数限制，以控制单个样本对参数更新的最大影响，从源头压缩整体敏感度，为差分隐私提供理论基础；其次，在裁剪后的梯度基础上注入满足差分隐私约束的高斯噪声，通过对聚合梯度扰动，增强模型对成员推断攻击的抵抗能力；最后，依据差分隐私理论设定隐私预算，并调节训练轮数与噪声强度，在隐私保护与翻译性能之间实现优化权衡。实验基于标准多语言翻译数据集，采用双语评估替代（BLEU）指标评估模型性能。消融实验进一步验证了三种技术的协同作用，结果表明，翻译质量下降控制在9%-28%之间，符合实际应用中的合理范围。在相同数据集上与其他机器翻译模型进行实验对比，尽管BLEU分数平均下降约5%-6%，但在保证翻译质量的同时，模型的隐私保护能力得到了有效提高。通过成员推断攻击实验，标准Transformer模型的攻击成功率为78.3%，而差分隐私mT5模型的攻击成功率降低至52.4%，进一步证明了本模型在隐私保护方面的优势。

Abstract: Aiming at the problem of data privacy protection in multilingual machine translation, A multilingual text-to-text transfer transformer (mT5) translation model incorporating a differential privacy mechanism is proposed, which protects user privacy while maintaining translation quality. First, gradient clipping is introduced during the model fine-tuning phase to limit the norm of each sample’s gradient, controlling the maximum influence of a single sample on parameter updates, thereby reducing overall sensitivity and providing a theoretical basis for differential privacy. Second, Gaussian noise that satisfies differential privacy constraints is injected based on the clipped gradients, and the model's resistance to member inference attacks is enhanced by perturbing the aggregated gradients. Finally, based on differential privacy theory, the privacy budget is set and training rounds and noise intensity are adjusted to achieve an optimal trade-off between privacy protection and translation performance. The experiments were conducted on a standard multilingual translation dataset, and the model performance was evaluated using the Bilingual Evaluation Understudy (BLEU) metric. The ablation experiment further verifies the synergy of the three technologies. The results show that the translation quality degradation is controlled between 9% and 28%, which is in line with the reasonable range in practical applications. Experimental comparison with other machine translation models on the same dataset shows that although the BLEU score drops by about 5% to 6% on average, the privacy protection ability of the model is effectively improved while ensuring the translation quality. Through the member inference attack experiment, the attack success rate of the standard Transformer model is 78.3%, while the attack success rate of the differential privacy mT5 model is reduced to 52.4%, further proving the advantage of this model in privacy protection.