The rapid shift toward digital job recruitment has unfortunately led to a rise in Online Recruitment Fraud (ORF), where cybercriminals exploit job seekers by advertising deceptive opportunities. Conventional machine learning (ML) techniques often fall short in detecting such fraud, primarily due to outdated data, reliance on manual feature extraction, and their inability to capture deep contextual meaning. In this work, we present a transformer-based deep learning framework leveraging Bidirectional Encoder Representations from Transformers (BERT) and its enhanced variant, RoBERTa, to robustly detect ORF. A unified dataset was curated by integrating three publicly accessible sources: the Fake Job Posting dataset, the Pakistan Job Market dataset, and the US Job Posting dataset. To mitigate class imbalance between real and fake job ads, we adopted oversampling methods such as Synthetic Minority Oversampling Technique (SMOTE) and Borderline-SMOTE (SMOBD). Contextual word embeddings generated by BERT and RoBERTa were further refined using a Two-Dimensional Convolutional Neural Network (2D CNN), enabling the model to learn both spatial and semantic features. The proposed model achieved a high F1-score of 0.93, demonstrating excellent precision-recall trade-off. Furthermore, a real-time fraud detection application was developed using the Flask web framework to ensure practical deployment and improved user trust on job platforms.