Universal Batch Steganalysis.

摘要

The overall theme of this project was to bring steganalysis into practical use, by proposing methods to identify a `guilty' user (of steganalysis) in large-scale datasets such as might be obtained by monitoring a corporate network or social network. Identifying guilty actors, rather than payload-carrying objects, is entirely novel in steganalysis, and we propose new methodologies to rank actors by their level of suspicion, without requiring any training data. We identified that modern so-called `rich' (large-dimensional) steganalysis features are not well-suited to unsupervised learning of this type, and developed novels ways to collapse large-dimensional data to reduce noise while retaining most of the evidence. These methods have been evaluated using large-scale experiments (in total over a million images tested in well over a billion combinations) on images crawled from genuine social networks. We also developed new implementations of existing steganalysis feature extractors, which were necessary for work on such a scale. We also examined a source of difficulty in all kinds of steganalysis: mismatch between actors caused by different cameras and post-processing. We proposed ways to mitigate such mismatch. Finally, we proposed a new method for attacking a single stego object by exhausting a secret key.