A Multilayered Risk-Aware Input Moderation Framework for Safe Large Language Models

SDGs: Primary SDG: SDG 9 - Industry, Innovation, and Infrastructure. | Secondary SDGs: SDG 3 - Good Health and Well-being, SDG 16 - Peace, Justice and Strong Institutions.

Abstract

Large Language Models (LLMs) have been used. transformed the way technology operates. However, their biggest weakness is the capability to generate psychologically dangerous information that may harm the vulnerable users. This is not a minor vice but one of the biggest barriers to their judicious adoption. Conventional safety approaches divided into fine-tuning and reinforcement learning (RLHF) are too complicated, non-transparent, expensive, and unable to estimate the fine balance between model utility and consumer wellbeing. The drawbacks of these being outweighed, a continuance has been made approaches, PTShield works as a preprocessing agent that guarantees psychologically safe interaction without affecting the model’s core parameters. This solution operates pre-inference by processing user script in real-time across value risk, goriest similar to suicidal ideation, toxic behavior, and de- pressive. language patterns. It offers real-times prescriptions based on the detections. It has been proved to be effectively demonstrated through. performance analyses showing superior test scores compared to traditional zero-shot, few- shot, and transfer learning approaches for risk content classification. The method prioritizes high-risk prompts, which reduces token wastage. Essentially, PTShield enables the development of psychologically sensitive and scalable safety systems for LLM deployments.