This chapter concludes this book on Watermarking Security and I hope that the reader will have a better view of the ins and outs of this domain. If watermarking security may look like a cat and mice game that is never ending, we can however state several important conclusions derived for the large variety of researches that have been conducted but there are also fascinating related problems that still need to be solved. The goal of this last chapter is to list the different outputs and open questions related to watermarking security but also general security of information forensics.

Conclusion 1 (connected with Chap. 3 ): there exits a methodology to analyze the security or increase the security of a watermarking scheme under a given scenario and given assumptions, this methodology takes into account the materials given to the adversary but also his goals and his power.

Conclusion 2 (connected with Chap. 3 ): several security proofs exist in watermarking security, such as the fact that an embedding is stego, subspace or key secure. These security proofs rely on non solved problems in signal processing, for example the fact that the separation of a mixture of Gaussian distributions of same variance is impossible, or the fact that if the host distribution is indistinguishable w.r.t. the watermarked one, the adversary can’t estimate the secret key in the WOA setup.

Problem 1: most of the proofs rely on the fact that the host distribution is perfectly known which is hardly true in practice since this distribution is a high dimensional function which is difficult to infer. We have here the same limitation than in steganography, perfect security is only possible if the host distribution of the medium is completely defined. One consequence of this problem is that it is difficult to prove that a scheme is stego or subspace secure once we use real multimedia contents: the problems of watermarking security and density estimation are intertwined.

Conclusion 3 (connected with Chap. 4 ): if the security class is acknowledged, it is possible to minimize the embedding distortion using the results of optimal transport theory when the host distributions are simple (e.g. separable).

Problem 2: for non trivial distributions, optimal transport theory may not provide solutions. For example the problem of performing optimal transport for any circular distributions is as far as I know still open.

Conclusion 4 (connected with Chap. 4 ): it is possible to measure the security of a scheme using either the information leakage of the secret key, or the effective key length as a maximal bound.

Problem 3: how can we derive an optimal key estimator? Here the optimality means that the average number of trials to have access to the watermarking channel is minimum.

Problem 4: once the security (i.e. the difficulty to break the watermarking system) is measured, how to compute the “secure capacity” of a watermarking channel? i.e. the maximum amount of information that can be securely transmitted? Theoretical frameworks such as the wiretap channel theory may provide partial solutions in this case [1, 2].

Problem 5: this problem is linked with problem 4, how to perform a joint optimization that takes into account both distortion, capacity and security? Here, partial solutions might be found by looking at the connection with other theoretical tracks such as adversarial learning [3], adversarial signal processing [4], and game theory. For example, there are important similarities between adversarial learning [3] and watermarking secure to Oracle attacks [5, 6].

Problem 6: the computation of the effective key length is a first step of an approach for computational security in watermarking. How to compute a similar measure when the adversary generates his own observations at will and looks at the output of the decoder/detector, such as for an oracle attack? what is the effective key length in this case?

By looking at this list of open questions, we can see that the field of watermarking security need still important research efforts, and the conclusions illustrate the fact that more than a decade after its birth, watermarking is not, as suggested Herley a nonsense [7] but on the contrary makes more sense than ever [8] and has to be considered with all its specificities.