Steganalysis based on transfer learning

Abstract

In practice, when the training set and testing set are mismatched, performance of steganalysis can not be guaranteed. The transfer learning aims at using the knowledge learned from one domain to help complete the learn-ing task in the new domain, and does not require the same distribution assumption. A more comprehensive review of mismatched steganography research status was made and the mismatch factors were analyzed. Methods on in-stance-based transfer learning were presented to solve the test mismatch problem during the steganography detections.

Keywords

• steganography;transfer learning;based on instance