Introduction
Reversible Watermarking (RW), or Reversible Data Hiding, is a critical technique in digital forensics and medical imaging where a secret message must be embedded into a digital cover without permanently altering the original data. The primary objective is to maximize embedding capacity (rate) while minimizing visual degradation (distortion). Traditionally, game-theoretic models in this field have focused on the conflict between a data hider and an active adversary. This paper shifts that focus toward a "hider vs. hider" dynamic, modeling the competition between multiple legitimate users for the limited "smooth" space available in a cover image.
Modeling Smoothness as a Finite Resource
The authors define the "smoothness" of a cover image using information entropy ($H$). In content-adaptive RW, data is most efficiently hidden in areas of low entropy. When multiple encoders—Alice and Bob—attempt to hide data in the same object, they essentially compete for these high-efficiency regions. The paper models the cover as a collection of binary sequences, where Alice embeds her payload first, followed by Bob.
Non-Cooperative Games and the "Last Mover" Advantage
In a non-cooperative setting, Alice and Bob act independently to maximize their individual payoffs (data capacity). This scenario creates a significant asymmetry. Because Alice embeds first, she must anticipate the distortion Bob will introduce. Bob's embedding acts as "noise" that Alice's system must be able to correct to maintain reversibility.
As shown in the figure above, Bob generally enjoys a "last mover" advantage. If Bob's distortion exceeds a specific threshold (e.g., $p_{max} = 0.75$ in the example), Alice's effective payload capacity can be completely negated. This highlights that in a competitive environment, the second embedder has more freedom to utilize the cover's resources at the expense of the first.
Equilibrium Strategies
To find a stable solution where neither player can improve their payoff by changing strategies alone, the authors derive Nash equilibria. For covers divided into multiple sections, the strategy involves allocating embedding intensity across different substrategies.
The equilibrium is typically reached when players hit their distortion bounds. The authors demonstrate that these equilibria can be efficiently calculated using standard optimization techniques like gradient descent or binary search, as the payoff functions remain concave.
Cooperative Games and Multi-layering
The research also explores cooperative scenarios where Alice and Bob share information, such as embedding keys. The mathematical analysis reveals that a cooperative game between two encoders is equivalent to a single encoder performing multi-layer data hiding. In this mode, the players act as a single unit to optimize the total capacity of the cover, effectively eliminating the competitive interference seen in the non-cooperative model.
Conclusion
This game-theoretic framework provides a new lens for understanding multi-user reversible watermarking. By treating image smoothness as a finite resource, the authors successfully quantify the trade-offs between competing encoders. Future work may extend this binary model to more complex grayscale or color images and explore the dynamics of games involving more than two players.
Summary updated