Application on LArge and heterogeneous images database of Steganalysis techniK for Advances "into the wild" – ALASKA

The present ALASKA project studies the detection of data hidden within digital images. This technique, referred to as steganography, can be leveraged by criminal rings or terrorist groups. Therefore, developing steganalysis tools, that is for hidden data detection, is crucial for national defense and intelligence forces.

While numerous research works addressed the problem of steganalysis over the last decades, a vast majority of them rely on an experimental academic framework which is very restrictive ; in fact, their application in practical setup has never been studied. There is almost no prior works dealing with hidden data detection for the purpose of uses by agencies operational. This is the main reason why, despite a substantial number of prior works from state of the art that relies on academic setup, it remains currently impossible to develop a practical detection tool for steganalysis dedicated to meet the need of intelligence operations.

The goals of present ALASKA project are towards filling the gap between research works and practical needs of intelligence agencies by strengthening the ties built between those two communities. The research works conduct under the umbrella of this project will focus on applications, in practical setup, of advances proposed in the field of steganalysis.
The overall aim is to design media analysis and steganalysis detection methods that can be used over images that can be intercepted or seized whithin digital investigations. To move forward in this practical and operational direction, the ALSKA project is divided into four works packages each dealing with one of the main issues, and yet seldom addressed until now, towards practical steganalysis.

The first one focuses on the design on steganalysis methods for highly heterogeneous image datasets ; two complementary approaches will be studied: the first one, referred to as targeted, seeks at gathering similar images within homogeneous groups that can be exploited for inspection ; the second one that, on the opposite, aims at leveraging this diversity for discriminating more accurately specific footprints of data hiding.

The second work package aims at developing recent advances using deep learning techniques for steganalysis in a practical framework ; the main challenge being to be able to reproduce the performance, very promising in terms of detection accuracy, of those approaches within a context closer to the operational one.

The third work package is devoted to the design of detection method associated with "probative value", especially with a bounded false alarm probability and quantitative results. It is indeed crucial, for allowing a large scale application, to ensure a very low false positive rate.

Eventually, the fourth and last work package deals with the organization of an international steganalysis contest with the goal, on the one hand, to motivate the research community to look into the direction of practical applications in operational context and, on the other hand, to benefit from contribution of the international challengers. This contest will be designed to highlight the main challenge in practical applications, especially those described in the three previous points. The progress made by the challengers will be analyzed and will be developed by the partners whenever appropriate.

The consortium of ALASKA project gathers three teams having complementary knowledges and skills filling all the scientific fields addressed and all having expertises widely acknowledged internationally.