Towards Secure Prompt Processing: A Unified Framework to Detect, Sanitize, and Prevent Adversarial Prompts via Dual-optimized Threshold-aware Learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

In the modern era, Large Language Models (LLMs) are a cornerstone in the creation of intelligent systems, boasting significant capabilities in various fields.Adversarial prompts continue to pose a significant threat to LLM security, as they exploit the lack of distinction between trusted and malicious prompts.Such weaknesses are extremely critical to the trustworthiness, safety, and reliability of real-world implements.To tackle this, this study introduces a sanitization and high-confidence inference for effective LLM defense through isolation and neutralization gateway (SHIELDING), which consists of linguistic feature extraction, sentiment-based analysis, and a Threshold-aware Random Forest (TRF) classifier.The TRF model is designed to include class-specific threshold optimized via Crested Porcupine Optimizer (CPO), allowing the model to adaptively make decisions.The experimental results indicate that the proposed method has good performance with an accuracy of 99%, a classification Attack Success Rate (ASR) of 0.656%, and a real ASR of 3.994% showing that it is an effective method to improve the security of LLM.