Fine-Tuned GPT Models for Covert Data Leakage Detection in UAV Networks

Abstract\u2014The unmanned aerial systems (UAS) in civilian andcommercial applications has introduced significant cybersecuritychallenges, particularly concerning data leakage risks duringdrone operations. This research presents a novel approach todrone data leakage classification using Generative Pre-trainedTransformer (GPT) models, specifically addressing the criticalneed for intelligent threat detection in drone communicationnetworks. We propose a comprehensive framework that leveragesInter-Arrival Time (IAT) features extracted from drone networktraffic to classify data leakage risk levels using fine-tuned GPTmodels. Our methodology involves preprocessing a dataset of48,020 samples with 14 IAT-based features, implementing bothzero-shot and fine-tuned GPT classification approaches, andconducting extensive comparative analysis with traditional ma-chine learning baselines. The experimental results demonstratethat our GPT-based fine-tuned model achieves 89.2% accuracy,representing a substantial 36.8% improvement over zero-shotperformance and competitive results compared to traditionalapproaches. The fine-tuned model exhibits superior performancemetrics with 88.85% precision, 89.2% recall, 89.02% F1-score,and 94.56% ROC AUC. This research contributes to the in-tersection of natural language processing and cybersecurity bydemonstrating the effectiveness of transformer-based models fornumerical feature classification in security-critical applications.The proposed framework provides a foundation for real-timedrone security monitoring and establishes new directions forapplying large language models to cybersecurity challenges inautonomous systems.