supervisors: Olli Simula

In the history of human communication, the concept and need for secrecy between the parties has always been present. One way of achieving it is to modify the message so that it is readable only by the receiver, as in cryptography for example. Hiding the message in an innocuous medium is another, called steganography. And the counterpart to steganography, that is, discovering whether a message is hidden in a specific medium, is called steganalysis. Other concerns also fall within the broad scope of the term steganalysis, such as estimating the message length for example (which is quantitative steganalysis). In this dissertation, the emphasis is put on classical steganalysis of images first — the mere detection of a modified image — for which a practical benchmark is proposed: the evaluation of a sufficient amount of samples to perform the steganalysis in a statistically significant manner, followed by feature selection for dimensionality reduction and interpretability. The fact that most of the features used in the classical steganalysis task have a physical meaning, regarding the image, lends itself to an introspection and analysis of the selected features for understanding the functioning and weaknesses of steganographic schemes. This approach is computationally demanding, both because of the feature selection and the size of the data in steganalysis problems. To address this issue, a fast and efficient machine learning model is proposed, the Optimally-Pruned Extreme Learning Machine (OP-ELM). It uses random projections in the framework of an Artificial Neural Network (precisely, a Single Layer Feedforward Network) along with a neuron selection strategy, to obtain robustness regarding irrelevant features, and achieves state of the art performances. The OP-ELM is also used in a novel approach at quantitative steganalysis (message length estimation). The re-embedding concept is proposed, which embeds a new known message in a suspicious image. By repeating this operation multiple times for varying sizes of the newly embedded message, it is possible to estimate the original message size used by the sender, along with a confidence interval on this value. An intrinsic property of the image, the inner difficulty, is also revealed thanks to the confidence interval width; this gives an important information about the reliability of the estimation on the original message size. keywords: Machine Learning, Steganography, Steganalysis, Extreme Learning Machine, Artificial Neural Networks, Feature Selection, Re-embedding. vii te l-0 07 37 35 3, v er si on 1 5 O ct 2 01 2