Optimized MS-SIDM: A Context-Aware Sensitive Information Detection Framework via LLM Reasoning and Non-Linear Risk Quantification

The exponential growth of multi-source data aggregation has exacerbated sensitive information leakage risks, challenging existing privacy compliance frameworks. Traditional detection methods, such as BERT-based NER, often suffer from ‘‘contextual blindness’’ and fail to address the ‘‘Mosaic Effect’’—where non-linear risk accumulation occurs across heterogeneous data streams. To address these deficiencies, we propose the Optimized Multi-Source Sensitive Information Detection Model (MS-SIDM). Our approach employs Qwen2.5-7B as a semantic backbone, fine-tuned via Low-Rank Adaptation (LoRA) on a taxonomy-guided adversarial dataset to distinguish authentic leaks from high-entropy traps. Furthermore, we introduce a Non-Linear Risk Quantification Mechanism integrating a Square Root Scale Factor (Mi) and a Density Compensation Factor (Di), alongside a Dual-Risk Perception mechanism (Eenv). This framework enables contextual threat assessment to dynamically calibrate risk scores against multi-source collusion. Experimental evaluations demonstrate that MS-SIDM significantly outperforms SOTA baselines (e.g., RoBERTa, Llama 2), achieving an F1-score of 97.46% and a Hard-Exact Match (EM) accuracy of 91.60% under adversarial stress. By synthesizing LLM-based semantic reasoning with a rigorous mathematical risk framework, this study demonstrates strong potential as a solution for asynchronous privacy monitoring and automated compliance auditing, subject to further validation in active production settings.