Truth Under Pressure: A Deep Learning-Based Lie Detection System for Online Lending Using Voice Stress and Response Latency

The rapid increase in defaults in the online lending industry highlights significant flaws in current debtor verification, which largely relies on static, preparable interviews, leading to high non-performing loans. Existing research is fragmented: while Large Language Models (LLMs) show promise in question generation, their application is confined to non-financial domains like education, and lie detection studies often analyze modalities in isolation. This study addresses this critical gap by proposing the first integrated AI-driven system for this context. We solve the problem in two parts: 1) A Llama 3 LLM is fine-tuned to generate dynamic, biodata-tailored questions, preventing the rehearsed answers that plague static interviews. 2) A novel multimodal deep learning model is developed to analyze the response, uniquely fusing vocal acoustic features and response latency—two key deception indicators that prior work has failed to combine. The Llama 3 model produced a low perplexity score (2-3), and the lie detection model achieved 70% testing accuracy with a 70.9% F1-Score. Despite signs of overfitting, this framework provides a novel, intelligent decision-support tool to reduce fraud and manage default risks more effectively.