Image Steganalysis Based on Transfer Learning
Abstract: This project is focusing on image steganalysis based on transfer learning. Current steganlaysis techniques are mainly based on heuristic feature extraction and supervised learning. These methods can achieve good detection results on specific but limited testing samples. Due to the cover source mismatch problem of steganalysis, traditional learning based staganalyzer are usually fail when dealing with the big data scenario, especially when facing unknown samples without label information. This is also due to that traditional surpervised learning based steganlysis methods don't have the adaptivity for unknown distribution, un-well trained classifier usually fails on new samples and cannot updated online. Our research is focus on steganalysis using transfer learning. Transfer learning is a new machine learning method that applies the knowledge from related but different domains to target domains. It relaxes the two basic assumptions in traditional machine learning: (1) the training (also referred as source domain) and test data (also referred target domain) follow the independent and identically distributed (i.i.d.) condition; (2) there are enough labeled samples to learn a good classification model, aiming to solve the problems that there are few or even not any labeled data in target domains. We also will design the transfer learning algorithms and proper scheme for metric learning as well as local learning. It is significant in both the theoretical and applicative perspectives for the development of image steganalysis.
Keywords: data hiding; steganalysis; machine learning; transfer learning
Contact:
DONG Jing
E-mail: jdong@nlpr.ia.ac.cn