AI Model Training Data Privacy Protection Scheme Based on Local Differential Privacy

In this paper, we present a novel AI model training data privacy protection scheme based on local differential privacy (LDP), aimed at safeguarding sensitive data in distributed environments. The method introduces dynamic noise adjustment to balance privacy and accuracy, optimizing model performance without compromising privacy. During the data preprocessing phase, the scheme applies standardization, handles missing values, and ensures format consistency. Noise is added dynamically to both the data and model gradients during training to further enhance privacy. The proposed method's scalability and computational efficiency were validated in large-scale, real-time AI applications, demonstrating significant reductions in overhead compared to centralized differential privacy techniques. Our results show that the dynamic adjustment of noise helps maintain high model accuracy while offering robust privacy guarantees, making the scheme ideal for use in distributed AI systems.