Evolving equilibrium in adaptive embedding and detection games

This project falls within the Mathematical sciences research area.

The project lies in the area of steganography. Here one party wishes to hide (embed) information in an apparently innocent cover, by making small changes, and their opponent wishes to detect whether information has been hidden. It is now established that some changes are more detectable than others, and literature exists to assign a "cost" to each possible change. This is known as "adaptive embedding".

If the embedder makes high cost changes only rarely, the detector can incorporate this information into their algorithm, exploiting the adaptivity of the embedder. Thus we have a detection game, between the embedder and detector. The aim of the project is to determine the equilibria of such games, by evolving them experimentally. This has potential impact in other areas where one party wishes to do something that is not detected.

The project involves implementation of a number of components, and experimental evaluation of the outcomes:
Obtaining real-world image data sets on which to experiment.
Implementing existing costs and adaptive embedding algorithms from the literature. This gives a parameter space for the embedder.
Finding a suitable parameter space for a detector, by modifying state-of-the-art steganography detection methods.
Using computational game theory methods (specific methods may include alpha-beta pruning and the Lemke-Howson algorithm) to evolve an equilibrium.
This programme can be repeated with different adaptive embedding, and with the hiding performed in different objects (in our application images from different sources), to draw conclusions about both parties optimal strategy. In this two-player game, the objective is determined by the accuracy of the detector, as measured empirically on real-world image hiding algorithms.
We do not anticipate new theoretical results arising from this work; it is experimental.

Novelty of methodology
There is very little existing research on the evolution of equilibria in steganography detection games, and no serious application of computation game theory. The proposed methodology is new. The project will investigate other research on (non-steganographic) detection games, in case similar problems have been attacked there.
There is a connection with a certain type of adaptive steganography (known as "adaptive steganography by oracle"), published recently. This is an example of an optimal counter-strategy on the part of the embedder. If successful, our work extends this to a minimax strategy for the detector.
Potential collaborators include Tomas Pevny - CTU Pargue / Cisco Inc, and Patrick Bas - CNRS Lille, who are also working on different aspects of adaptive steganography.