Enhanced Online Recruitment Fraud Detection Using Transform- Based Deep Learning Model

Abstract: Online recruitment platforms have significantly transformed the job search process by improving connectivity between employers and job seekers. However, this rapid growth has also increased the vulnerability of such platforms to fraudulent job postings designed to deceive applicants through misleading offers, unrealistic benefits, and unauthorized payment requests. Recruitment fraud has emerged as a serious cybersecurity and social engineering concern, particularly affecting students, recent graduates, and unemployed individuals. Traditional detection approaches based on rule-based filtering techniques and classical machine learning models such as Logistic Regression, Support Vector Machines, and Random Forests demonstrate limited adaptability to evolving fraud patterns and depend heavily on manually engineered features. These approaches often fail to capture contextual and semantic relationships within unstructured job descriptions, resulting in reduced detection accuracy and poor generalization. To overcome these limitations, a Transformer-based deep learning architecture using DistilBERT is employed to automatically classify job postings as genuine or fraudulent. The methodology incorporates an end-to-end pipeline including text cleaning, normalization, tokenization, contextual feature extraction, and transformer-based classification using self-attention mechanisms. The model effectively captures long-range dependencies and linguistic indicators commonly associated with fraudulent postings, such as urgency-driven language, vague job descriptions, and payment-related requests. Experimental evaluation conducted on a publicly available dataset containing 17,880 job postings demonstrates strong discriminative performance under class-imbalanced conditions, achieving high accuracy, precision, recall, and AUC-ROC scores. High precision ensures minimal false alarms for legitimate postings, while the strong AUC-ROC performance indicates reliable classification capability across different decision thresholds.