Decision-Theoretic Steganography

• Decision-Theoretic Steganographic Formalism is a unified framework that leverages decision, game, and utility theory to define and secure covert communication.
• It integrates indistinguishability games, constrained optimization, and risk analysis to quantify the operational impact and detection risk of hidden messages.
• The framework employs strategies from CMDP and multi-agent reinforcement learning to optimize embedding rates while ensuring imperceptibility and robust security.

Decision-theoretic steganographic formalism provides a unified mathematical foundation for the quantification, optimization, and empirical assessment of steganography through the lens of decision theory, game theory, and utility-based information measurement. Unlike purely statistical approaches to steganographic security—where secrecy is guaranteed by indistinguishability from cover distributions—decision-theoretic frameworks focus on the operational effects of hidden information, the consequences of agent interaction, and adversarial risk, enabling a rigorous analysis in both classical and modern (e.g., LLM-reasoning) settings. This perspective encompasses indistinguishability games, strategic utility maximization, constrained optimization, and utilitarian measures of information asymmetry.

1. Indistinguishability and Decision-Theoretic Security Foundations

Steganographic security, in its foundational decision-theoretic formalism, is characterized by indistinguishability games between a stegosystem and a passive observer, or "warden." Formally, a stegosystem is a pair of probabilistic polynomial-time algorithms, SE (encoder) and SD (decoder), interacting with a channel 𝒞 that emits sequences of cover objects. Security is defined by the inability of any polynomial-time warden to distinguish between a stegotext (produced by SE embedding a hidden message) and a genuine cover sequence sampled from 𝒞, formalized by the advantage metric: $$\mathrm{Adv}_{S,\mathcal{C}}^W(k) = \left| \Pr[W^{M,O_0} \Rightarrow 1] - \Pr[W^{M,O_1} \Rightarrow 1] \right|$$ where $O_0$ produces stegotexts and $O_1$ samples pure covers (Alston, 2017). The standard of (t, q, ℓ)-security requires this advantage to be negligible for all polynomial resources.

Security reductions typically relate the indistinguishability of the stegotext from cover to the cryptographic hardness of underlying primitives (e.g., PRFs), ensuring that any warden able to distinguish also breaks the pseudorandomness, establishing universal security in practical constructions such as block-cipher IV-steganography (Alston, 2017).

2. Extensions Beyond Syntactic Models: Semantic and Utilitarian Views

Traditional formalisms are often syntactically constrained, focusing on fixed-length blocks and explicit sampling oracles. Generalization to broader domains—multimedia, natural language, and protocol stacks—favors an abstract view: any stegotext sequence must be distributed indistinguishably from the channel's output. This semantic shift allows for flexible security definitions and applicability across complex message spaces (Alston, 2017).

Recent advances introduce a utilitarian perspective: rather than security solely as statistical indistinguishability, steganographic risk is conceptualized in terms of operational effects on utility. The central insight is that steganography creates information asymmetry—agents with the decoding key can extract actionable knowledge, while others cannot. Starting from the framework of ∨-information, the maximal expected utility achievable by an agent with or without the hidden signal is compared, forming the "steganographic gap": $$\Delta_{\mathrm{steg}}(Z) = I_{\mathrm{rec}} - I_{\mathrm{sen}}$$ where $I_{\mathrm{rec}}$ and $I_{\mathrm{sen}}$ are gains in maximal utility for Receiver and Sentinel families, respectively. This measure subsumes classical total variation-based secrecy and directly quantifies the practical impact of hidden-information transfer (Anwar et al., 26 Feb 2026).

3. Game-Theoretic and Risk-Analytic Steganography

Decision-theoretic steganographic formalism extends naturally to strategic, adversarial environments, where defenders and attackers select actions to optimize their objectives under uncertainty. This is formalized as non-cooperative games, often with a one-shot, simultaneous-move structure:

• Defender ($U$) chooses whether or not to employ steganography.
• Adversary ($A$) chooses whether or not to attempt detection.

Utilities incorporate monetary and operational costs (cost of steganography, loss on leak, gain from safe transmission, attacker's cost/reward). Nash equilibrium is derived by imposing mutual indifference conditions, yielding closed-form mixed strategies: $$p^* = \frac{G - B^{U}_{\text{hide}} + C_s}{G + L_{\text{leak}}}, \qquad q^* = \frac{R - C_d}{\beta_U R}$$ with adversary advantage $A_{\text{adv}}^* = p^* \beta_U q^*$ quantifying detection risk (Omego et al., 2024).

Monte Carlo simulations with parameter uncertainties (impact factor, success rates, costs) provide empirical risk distributions. Sensitivity analysis assesses the impact of parameter shifts, and explicit decision rules determine when embedding is justified by utility gains or bounded risk (Omego et al., 2024).

4. Steganographic Gap and Generalised V-Information

The decision-theoretic lens reframes steganography detection as identifying an actionable information gap between agents with and without decoding capabilities. Generalised V-information operationalizes this by measuring, for a set of observations ($O_0$0), signals ($O_0$1), and tasks ($O_0$2), the maximal utility gap achievable by decoders in specified families (e.g., Receivers with the key, Sentinels without): $O_0$3 The normalized steganographic gap,

$O_0$4

ranges from 0 (Sentinel matches Receiver) to 1 (total asymmetry). Empirically, this methodology is validated across simple ciphers, RL-trained codebooks, and real LLM backdoor attacks. Increasing cipher complexity or key exclusivity raises the steganographic gap; making the Sentinel more powerful or applying paraphrasing closes it (Anwar et al., 26 Feb 2026).

5. Sequential, Constrained Optimization Formalisms

In sequential domains, steganographic embedding is formalized as a constrained Markov Decision Process (CMDP), maximizing embedding rate subject to imperceptibility constraints. States represent contexts (e.g., token sequences), actions specify modified emission distributions, and the reward is per-symbol embedding entropy. The cost, total variation between emission and baseline, is globally bounded: $O_0$5 where $O_0$6 quantifies detectability (e.g., TV divergence) (Huang et al., 3 Feb 2025).

The core result is that, for finite state-action CMDPs, optimal policies are deterministic and exhibit "water-filling" structure: the policy prioritizes adjustment in states with lowest cover entropy, subject to the global imperceptibility budget. This contrasts with local greedy or randomized embedding, leading to improved global efficiency (Huang et al., 3 Feb 2025).

6. Multi-Agent and RL-Driven Steganography

Steganographic formalism generalizes to multi-agent environments where agents (e.g., “Alice” and "Bob") embed and decode hidden messages through coordinated actions in complex environments, modeled as stochastic games or MDPs. Here, encoder policies are trained to optimize both the overt task and covert communication utility, while observer (decoder) policies attempt to infer the hidden message from trajectories. Additionally, adversarial agents may add noise or attempt detection.

Key performance metrics include:

• Distortion: Discrepancy between stego and cover trajectories.
• Capacity: Entropy of hidden messages per episode.
• Secrecy: Statistical divergence (KL, TV) between stego and cover distributions.
• Robustness: Decoding accuracy under adversarial perturbation (Chang et al., 2024).

Multi-agent RL algorithms interleave Q-learning for encoder agents and supervised classification for observers, with equilibria determined by the tradeoff between reward for cooperation (effective hiding) and the risk of exposure. Laboratory results demonstrate that significant hidden-capacity can be achieved with near-baseline overt performance and high secrecy (Chang et al., 2024).

7. Comparative Table of Decision-Theoretic Steganographic Formalisms

Reference	Core Formalism	Key Application Domain
(Alston, 2017)	Indistinguishability experiment, PRF security	Classical stegosystems
(Omego et al., 2024)	Game-theoretic utility, Nash equilibrium, risk	Risk analysis and adversarial IR
(Huang et al., 3 Feb 2025)	CMDP optimization, TV-constrained entropy	LLM-based linguistic steganography
(Chang et al., 2024)	Multi-agent RL, game-theoretic, RL equilibrium	Action/trajectory steganography
(Anwar et al., 26 Feb 2026)	∨-information, steganographic gap, utilitarian	LLM monitoring & behavioral audit

This comparative summary contextualizes key advances and domains for each formalism, highlighting the evolution from indistinguishability-centric security to utilitarian, game-theoretic, and RL-constrained optimization perspectives.

8. Significance and Open Problems

Decision-theoretic steganographic formalism establishes a rigorous, flexible basis for both analyzing and engineering secure stegosystems under pragmatic constraints and adversarial dynamics. Its significance lies in:

• Enabling provable security reductions to standard cryptographic primitives in classical domains.
• Providing operational, task-dependent detection and quantification methods in scenarios lacking explicit reference distributions (e.g., LLMs).
• Supporting optimization of embedding strategies under global imperceptibility or risk budgets.
• Integrating utility, information, and strategic equilibrium into a unified analysis.

Open challenges include extending these frameworks to dynamic multi-agent networks, adaptive adversaries, and continuous/semantic domains where cover distributions are nontrivial or evolving, and ensuring that utility-gap measures robustly capture practical stealth and risk in real-world deployments.