Reconnaissance is the first step in the Cyber Kill Chain (CKC) model, which aims to gather se-curity information about a target. In addition to using network and vulnerability probing tools, sophisticated attackers also apply social engineering (SE) techniques to exploit dupes to gain access to sensitive/private data of the target, as human beings are often the weakest link in security. Therefore, this paper proposes a “honey imprint” enabled approach which takes ad-vantage of Natural Language Processing (NLP) for detecting SE attacks, Generative Adversarial Networks (GAN) for generating decoy from original sensitive information, and steganography for imprinting the honey water-mark. The purpose of honey-imprint is to protect the sensitive infor-mation contained in the original file while leaving a covert imprint on the honey file for identifying the malicious user. With this, we can further capture malicious interactions (using honey-im-printed data) by the honeypot system. We implement a prototype to verify the design, and the experimental results show that the method is valid and effective.