Abstract
Traditional watermarking typically focuses on the detection of discrete messages. This paper introduces "soft watermarking," a framework where the hidden information is a continuous signal rather than a binary ID. We model the interaction between an encoder-decoder team and a malicious attacker as a zero-sum game within a scalar Gaussian environment. By minimizing the Mean Squared Error (MSE) of the watermark estimation under specific power constraints, we derive optimal linear strategies for both parties. Our results characterize the transition between trivial and non-trivial communication regimes and prove that the optimal attack strategy against a linear encoder is Gaussian-affine.
Introduction
In the field of digital security, watermarking is often treated as a binary decision problem: a copyright ID is either present or absent. However, many practical scenarios require hiding continuous data—such as embedding a low-resolution preview within a high-resolution image. This "soft watermarking" paradigm shifts the problem from joint source-channel coding to joint source-source coding. This paper provides a rigorous mathematical foundation for this approach, analyzing the strategic "tug-of-war" between a data-hider and an intelligent adversary using game theory.
System Model and Problem Formulation
The soft watermarking system is composed of three primary actors: the encoder, the attacker, and the decoder. The goal of the encoder-decoder team is to transmit a watermark signal while minimizing distortion, while the attacker seeks to maximize the estimation error of that watermark.
As illustrated in Figure 1, the process involves two independent Gaussian random variables: the watermark $X \sim \mathcal{N}(0, \sigma_x^2)$ and the host signal $S \sim \mathcal{N}(0, \sigma_s^2)$.
- Encoder: Combines $X$ and $S$ into a transmitted signal $U = g(X, S)$. The encoder must respect a distortion budget $P_E$, ensuring that the modification to the host signal is limited: $E[(U-S)^2] \leq P_E$.
- Attacker: Intercepts $U$ and produces a corrupted signal $Y = \mu(U)$. The attacker aims to maximize the final estimation error while staying within its own power budget $P_A$, such that $E[(Y-U)^2] \leq P_A$.
- Decoder: Receives the attacked signal $Y$ and produces an estimate $\hat{X} = h(Y)$. The objective of the decoder is to minimize the Mean Squared Error (MSE), defined as $J = E[(X - \hat{X})^2]$.
Game-Theoretic Analysis
The interaction is modeled as a zero-sum game. The authors assume a scalar Gaussian environment and focus on linear strategies for the encoder, where $U = \alpha X + \beta S$. A major contribution of this work is the proof that for any linear encoder, the most effective strategy for an attacker is to behave as a Gaussian-affine channel: $Y = \kappa U + Z$, where $Z$ is additive Gaussian noise. This finding allows the complex functional optimization problem to be reduced to a more manageable parameter optimization over the weights $\alpha, \beta, \kappa$ and the noise variance $\sigma_z^2$.
Key Insights and Results
The analysis identifies two distinct operational regimes based on the system's power constraints and the host signal's variance:
- Non-Trivial Regime: When the attacker's budget is limited relative to the encoder's capabilities, the encoder can successfully embed the watermark. The optimal strategy involves a specific weighting of the watermark and host signal, which the authors derive using a depressed cubic polynomial.
- Trivial Regime: If the attacker's power $P_A$ is sufficiently large, they can effectively "erase" the watermark's influence on the signal. In this state, the received signal $Y$ becomes independent of $X$, and the MSE reaches its maximum value, $\sigma_x^2$.
Interestingly, the transition between these regimes does not depend on the power of the watermark itself ($\sigma_x^2$), but rather on the relationship between the host signal power and the two distortion constraints ($P_E$ and $P_A$). In the "Large Signal Case," where the host signal is significantly stronger than the noise and budgets, the optimal cost $J$ simplifies to a clean ratio involving the attacker's power and the host signal's power.
Conclusion
This research establishes a formal framework for soft watermarking under adversarial conditions. By proving the optimality of Gaussian-affine attacks against linear encoders, the authors provide closed-form solutions for strategic data hiding. While this study focuses on the scalar case, it lays the groundwork for future extensions into high-dimensional vector spaces, which are more representative of real-world multimedia data like video and audio.
Summary updated