A novel privacy-preserving large language model integrating trust-weighted and ethical gradient masking

Large language models face a critical privacy utility tradeoff because training data contains personally identifiable information vulnerable to extraction attacks, while existing differential privacy methods significantly degrade model performance. This work proposes the trust-weighted and ethical gradient model, a novel privacy-preserving model that addresses challenges through three integrated mechanisms. First, trust-weighted memory employs per-sample weighting based on data sensitivity to reduce memorization of sensitive information while preserving learning signals from non-sensitive data. Second, ethical gradient masking combines per-sample clipping and Gaussian noise with sensitivity-based soft masking under a tracked ( ϵ , δ ) privacy budget, while maintaining utility. Third, the ethical boundary layer classifies and mitigates risky outputs before delivery, providing a final defense against privacy leakage. A distinctive feature of the proposed approach is the integration of computational interpretations of trustworthiness, privacy, and non-intrusion into artificial intelligence privacy systems. Extensive experiments on the Pile dataset demonstrate that trust-weighted and ethical gradient models achieve a reduction in personally identifiable information leakage while maintaining baseline utility, outperforming baseline differential privacy methods in this evaluation. The model exhibits strong resistance against membership inference, extraction attacks, and linkage attacks for privacy-preserving large language model training. The selective masking improves utility compared to uniform differential privacy under the same privacy budget.