Reviewed by
Adam Wierzbicki
Polish-Japanese Institute of Information Technology
The modeling approach used in the book is classical game theory. The book begins with a good introduction that covers the basic concepts of game theory. The book focuses on noncooperative game theory, and supplements theoretical analysis of equilibriums with simulation results. This makes the title somewhat too broad. Game theoretic models for sharing of scalable media would have been better. Social networks, on the other hand, are hardly used in the book, except for a few chapters where the simplest possible topologies (like a star or a clique). Real social network topologies derived from empirical data are not used in the experiments described in the book. Neither are valid strategies for attacking colluder social networks that are based on social network structure.
The research method used throughout the book is classical game theory and simulation. Classical noncooperative game theory is the basis of most models considered in the book. This is natural considering that the basic research problem is the sharing of media that contain watermarks. These watermarks must be removed by sharing users who wish to use the copyrighted media without payment. The basic model in the book is a two-person game played by a colluder (modeling all colluding users) and a detector (trying to catch the colluder using the watermark). The game is a Stackelberg (leader-follower) noncooperative game. A second model is a "multiuser collusion game" where the two players represent users who receive a low quality version of the media, and users who receive a high quality version. A chapter of the book devoted to cooperation stimulation in Peer-to-Peer video streaming also uses noncooperative, two player games.
The book focuses on analyzing the game-theoretic models by computing Nash equilibriums and by equilibrium refinement, especially finding Pareto-optimal Nash equilibriums and applying the concept of equal-risk fairness. This approach is complemented by simulation of more complex versions of the models.
Although theoretically valid, the results obtained in the book concern mostly Nash equilibriums. Since the 1990s, behavioral game theory has repeatedly established that Nash equilibriums do not produce a good fit with the behavior of users in experimental games. There are several reasons for these discrepancies, but a widely accepted view is that the Nash equilibrium can no longer be considered as a default solution of game-theoretic models. What could be used instead is a Quantal Response Equilibrium (QRE), where players do not choose the best action with probability one (as in the Nash equilibrium) (e.g., Camerer 2003) In the QRE, players form beliefs about what others will do when they choose an action, and then compute expected payoffs given those beliefs.
Research in behavioral game theory has demonstrated the ability of QRE to explain departures from Nash equilibrium behavior in several games. In other words, QRE equilibriums are able to predict real player behavior, while Nash equilibriums are not generally useful for this purpose. As stressed by Camerer, "Quantal response equilibrium (QRE), a statistical generalization of Nash, almost always explains the direction of deviations from Nash and should replace Nash as the static benchmark to which other models are routinely compared" (Camerer 2003).
QRE equilibriums are quite difficult to compute (finding the optimal mixed strategy is equivalent to optimizing a nonlinear and non-convex function). However, they are much simpler to simulate. As most game-theoretic studies employ simulation for verifying more complex hypotheses, they could also be extended with a simulation that would calculate QRE equilibriums. The found QRE equilibriums could be then compared to mixed-strategy Nash equilibriums. If the two kinds of equilibriums are dissimilar, the QRE equilibrium should be used instead of the Nash equilibrium in subsequent analysis.
In chapter 8, the authors use the Nash bargaining solution to study fair collusion. This is another classical game-theoretic concept that should be approached with extraordinary caution by the contemporary modeler. The Nash bargaining solution relies on several restricting assumptions that do not necessarily apply in practice. The Kalai-Smorodinsky bargaining solution is a different approach that can be considered as more realistic.
Despite the use of simulation to study more complex model variants, the book contains no scenarios that would apply learning-based or heuristic strategies of the users. User behavior in the book is consistently modeled using the standard game-theoretic assumptions of strategic and selfish behavior with full information about the game payoffs. Such assumptions are not likely to hold in realistic scenarios of media sharing. Standard game-theoretic assumptions can be relaxed in simulations by allowing players to use strategies that rely on learning or evolutional heuristics.
Part V of the book concerns the effect of social network structures on collusions of media-sharing users. This part of the book contains a comparison of a centralized and decentralized network topology. The centralized topology contains a trusted ringleader, while the decentralized one requires the use of a special algorithm for collusion of peers in a distributed social network where each peer can contact all other peers.
The experiments described in the book do not consider the dynamics of social networks, particularly attack/defense strategies that are based on social network structure. Such strategies can include: attacking most connected users, defending by forming cliques, attacking based on centrality, defending based on delegation and cliques. These strategies have been shown to be effective in many studies (e.g., Anderson et al. 2007).
The book also deals with fairness, which is understood as an equality of risk (for colluders). Risks are borne by colluders who share their versions of the received media, because they can be detected. Equality of risks is an intuitive definition of fairness. However, it excludes the possibility of trading off risk equality against overall efficiency (minimization of risk for all colluders). Using a model where users may have varying risks would raise questions of coalition formation, requiring use of cooperative game theory. In the non-cooperative case, identifying fairness with equality of risk is quite sensible.
To sum up, the book has a good structure and is well written, although is restricted to an advanced audience, as computing game-theoretic equilibriums requires mathematics that may be too hard for an undergraduate reader. The strong point of the book is a systematic study of using digital watermarking for controlling dissemination of information by groups of media sharing users. The book demonstrates the effectiveness of using a consistent, simple theoretical approach such as non-cooperative game theory. Extending and modifying the used game-theoretic models has enabled the authors to study an impressive variety of issues. Finally, the results can be of practical importance for designers of protection mechanisms based on watermarks. However, it is worth noting that other methods exists that trace back to data provenance techniques (e.g., Papadimitriou and Garcia-Molina 2011).
Also, the validity of the results in the book may be questioned if the standard game-theoretic assumptions that are the basis of most results do not apply. This may concern realistic behavior of media sharing users. Considering attack strategies that are based on social network structure would also be required to better understand the dynamics of collusion in sharing of watermarked media.
CAMERER C, (2003) Behavioral Game Theory, Princeton: Princeton University Press
PAPADIMITRIOU P and Garcia-Molina, H (2011) Data Leakage Detection. IEEE Transactions on Knowledge and Data Engineering, 23, 1, pp. 51-63