<p>Cyber deception has emerged as a powerful complementary strategy to enhance traditional security against dynamic and sophisticated attacks. Recent research focuses on generating and deploying deceptive documents to protect the document against intellectual property (IP) theft by misleading the attacker towards fake documents. A set of methods has been proposed to generate decoy documents against IP theft. However, previous works have not focused on testing the robustness of generated fakes against attackers’ digital forensic processes. Attackers often employ forensic techniques to verify stolen data for credibility in the darknet market. In this work, we focus on generating realistic decoy documents and employ unsupervised clustering techniques to distinguish between original and counterfeit documents, thereby assessing the robustness of the generated fakes. We introduced a framework BFDGBMI (Believable Fake Document Generation using BERT-based Masked Infilling) for generating fake documents using a BERT-based masked infilling approach. It enhances the process of counterfeit document generation by solving the limitations of the GPT-based fake document infilling (FDI) technique. We have fine-tuned the BERT-based uncased model on the collected datasets and generated fake documents for each test data entry. Thereafter, tested the robustness of the deployed fake documents using a set of unsupervised clustering techniques. Additionally, we used the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) and coherence scores to evaluate the change in the machine generated fakes. To assess the believability of the generated fakes, a human evaluation study was conducted among students. The generated fake documents are indistinguishable from the original documents.</p>