Covert Semantic Communication Framework

• Covert semantic communication is a framework that embeds high-level semantic content into signals while maintaining strict covertness via hypothesis testing and square-root scaling laws.
• Architectures integrate cross-layer coding, steganography, and RL-driven scheduling to optimize power/rate allocation and obfuscate transmission from adversaries.
• Security evaluations use metrics like BLEU scores, detection AUC, and rate-distortion bounds to assess the trade-offs between embedding capacity and computational overhead.

Covert Semantic Communication Framework

Covert semantic communication frameworks enable entities to embed, transmit, recover, and analyze high-level semantic content with guarantees of covertness: adversaries (“wardens”) should have negligible probability of reliably determining whether and what semantic information has been communicated. Modern frameworks operate across the intersection of information theory, deep learning, optimization, multi-agent systems, and security, supporting applications in regulated wireless networks, military scenarios, strategic language games, and agent collaboration under adversarial scrutiny.

1. Foundational Principles and Mathematical Models

Covertness is formalized as a hypothesis testing problem between an innocent (“cover”) and an active communication scenario, with the warden’s detection probability bounded by a parameter $\epsilon$ or the divergence between observed distributions. The semantic channel is modeled jointly with a source process (e.g., image, text, agent utterance) and a channel (noisy, broadcast, or multi-hop), with the objective to transmit an intended meaning or achieve a downstream task, all while maintaining stealth.

Information-theoretically, covert semantic communication is subject to strong scaling laws. If a transmitter observes a semantic sequence $U^k$ and a channel is used $n$ times, then under stringent covertness constraints (e.g., KL divergence $$\mathbb{D}(\widehat Q^n\;\|\;\Gamma_{Z|X}^{\otimes n}(\cdot\mid0^n))\le\delta_n$$), the maximal extractable semantic bits $k$ scales only as $O(\sqrt{n})$, a square-root law analogous to classical covert channel capacity (Bounhar et al., 2024). For reliable communication, the achievable distortion $D$ satisfies $R(D) \le \gamma C_{\rm covert}$, where $R(D)$ is the semantic rate-distortion function and $C_{\rm covert}$ is the channel’s covert capacity.

Optimization problems arise both in power/rate allocation—balancing semantic fidelity and covertness—and in joint selection of coding, transmission slots, and path selection under multi-agent settings or adversarial uncertainty (Liu et al., 19 Jan 2025).

2. Architectures and System Designs

Frameworks span diverse architectures tailored to application and threat models:

• Cross-layer semantic frameworks: Layered architectures integrate application-layer semantic coding, network-layer path and relay selection, and physical-layer embedding of covert symbols within noise-shaping or waveform modification (Liu et al., 19 Jan 2025). Semantic encoders (e.g., transformers, VAEs) extract and compress key features, while generative AI components synthesize statistically-matching cover traffic.
• Steganographic and coverless designs: Advanced approaches avoid explicit cover objects, generating plausible reference modalities conditioned on public semantic content and user tokens. For example, AgentSemSteCom utilizes diffusion generators to synthesize reference images and encodes secrets directly into the semantic latent space, eliminating the need for a pre-selected cover and minimizing the risk of semantic key leakage (Meng et al., 23 Jan 2026).
• Adaptive multi-agent and dual-path networks: Frameworks such as CoMet (Xu et al., 23 May 2025) and the adaptive dual-path SemCom model (Yu et al., 5 May 2026) enable semantic-level intrinsic embedding (e.g., metaphors in language games or feature-aligned neural representations), with block-wise adaptive selection and contrastive alignment to make covert and overt paths statistically indistinguishable.
• Contract-theoretic and entropy control protocols: For distributed, incentive-misaligned agents (e.g., UAVs in LAENets), contract-based mechanisms regulate abstraction levels and semantic entropy to maximize overall task value while provably limiting detection risk. Soft actor-critic and diffusion-regularized RL algorithms are leveraged for robust policy learning under prospect-theory-based utility (Liu et al., 2 Mar 2026).
• Slot-optimal RL-driven scheduling: In time-slotted scenarios with jamming-based cover refinement, prioritized sampling techniques (e.g., PS-TD3) optimize both the timing and power allocation of semantics given incomplete knowledge of the jammer’s pattern and the attacker’s monitoring budget (Zhang et al., 11 Aug 2025).

3. Core Methodologies and Algorithms

A set of algorithmic components recur across frameworks:

• Semantic encoder/decoder design: Compresses and reconstructs data at the meaning level, subject to rate-distortion-covertness constraints. Joint source-channel neural codecs (e.g., JSCC, transformer-based, INN-based) are prevalent (Tang et al., 29 Mar 2025).
• Covert embedding and steganography: Hiding modules (e.g., invertible neural networks, coupling blocks, diffusion-based perturbations) ensure that transmitted representations match the distribution of innocuous traffic, evading statistical and machine-learning–based detectors, sometimes via explicit loss regularization on embedding and null hypothesis similarity (Cao et al., 23 Dec 2025).
• Experience-driven or RL-based adaptation: Agents refine strategies (e.g., metaphor use, slot selection) using in-game or online feedback, experience pools, or prioritized replay, exploiting soft actor-critic, diffusion policy, or delayed double-Q updates to traverse high-dimensional non-convex solution spaces (Xu et al., 23 May 2025, Zhang et al., 11 Aug 2025, Liu et al., 2 Mar 2026).
• Contrastive and alignment losses: To obfuscate the presence of covert information, contrastive representation alignment brings the joint “stego+task” path distributions close to those of explicit-only transmissions (Yu et al., 5 May 2026).
• Prefix-independent decoding: In decentralized or memory-augmented agent networks, prefix-independent decoding rules permit robust steganographic extraction even as agent contexts diverge, decoupling semantic generation from covert extraction (Wu et al., 9 Apr 2026).
• Regularized diffusion optimization: Contract design and policy learning benefit from diffusion-based entropy regularization, which stabilizes updates and drives policies toward expert-like sampling (Liu et al., 2 Mar 2026, Liu et al., 19 Jan 2025).

4. Security and Performance Evaluation

Covert semantic frameworks are subject to stringent evaluation under varied attack models:

• Adversarial detection and eavesdropping: Wardens employ model inversion, semantic-level discriminators, and adversarially trained deep learning models to detect statistical anomalies or extract private semantics. Advanced frameworks enforce computational indistinguishability (BERT/ML classifier accuracy ≈50%) and ensure that even strong (glass-box and GenAI) attacks reduce to random-guessing on hidden data (Meng et al., 23 Jan 2026, Tang et al., 29 Mar 2025, Cao et al., 23 Dec 2025).
• Fidelity/stealth trade-off metrics: Metrics include semantic quality (BLEU, mIoU, GNT, PSNR, SSIM), privacy leakage (mutual information, direct similarity between eavesdropper’s output and source, eavesdropping success rates), statistical distinguishability (entropy per token, detection AUC), and effective covert capacity (EIC, bits/1k tokens) (Wu et al., 9 Apr 2026, Cao et al., 23 Dec 2025).
• Scaling laws and task robustness: Empirical results confirm that, under covert constraints, the number of semantic features or bits must scale sub-linearly ($U^k$0) with resources to prevent detection, with significant performance loss when pushing toward the non-covert regime (Bounhar et al., 2024).
• Sample performance outcomes: On military IoT cross-layer networks, diffusion-empowered RL yields 8–9% higher semantic accuracy and covert success rates than standard SAC or PPO (Liu et al., 19 Jan 2025). In dual-path SemCom, attacker detection is reduced to ≈52.8%—near the random limit—while maintaining high public and covert semantic accuracy (Yu et al., 5 May 2026). INN-based signal steganography eliminates facial privacy eavesdropping success (FPESR drops from >80% to 0%) (Tang et al., 29 Mar 2025).

5. Application Domains and Practical Realizations

Key domains for covert semantic frameworks include:

• Regulation-aware wireless networks: UAV-launched LAENets must comply with low-probability-of-detection mandates, dynamically adjusting semantic entropy, modulation, and abstraction levels to trade off task performance against covertness (Liu et al., 2 Mar 2026).
• Multi-agent interactive games: Agents employ hypothesis-based semantic alignment (e.g., metaphor, irony) for semantic evasion, coordinated behavior, and adversarial misdirection in both competitive and mixed settings (Xu et al., 23 May 2025).
• Cloud–Edge IoT and military networks: Cross-layer semantic security ensures end-to-end confidentiality, strategic channel selection, and robust adaptation to evolving adversary detection strategies (Liu et al., 19 Jan 2025).
• Video and image semantic communications: Deep hiding modules and randomized embedding obfuscate the presence of secret content without degrading overall link quality. Covert video frameworks operate at the latent semantic level, ensuring that both overt and secret streams are recoverable only by credentialed parties (Cao et al., 23 Dec 2025, Meng et al., 23 Jan 2026).
• Cognitive agent networks: Prefix-independent encoding/decoding and steganographic partitioning overcome cognitive asymmetry in agent communication, supporting robust peer authentication and control signaling (Wu et al., 9 Apr 2026).

6. Limitations, Open Challenges, and Research Horizons

While covert semantic communication frameworks exhibit strong security and fidelity, limitations persist:

• Embedding rate limitations: Raw covert embedding rates are fundamentally limited by covertness constraints and channel conditions, with square-root scaling in many scenarios (Bounhar et al., 2024, Wu et al., 9 Apr 2026).
• Computational overhead: Generative models and joint RL optimization introduce significant computational and latency burdens; strategies such as knowledge distillation and quantization are proposed for mitigation (Liu et al., 19 Jan 2025).
• Jamming and power allocation trade-offs: Balancing friendly jamming to mask transmissions without overly degrading legitimate semantic decoding remains an open tension, exacerbated under energy constraints and incomplete knowledge of interference patterns (Zhang et al., 11 Aug 2025, Liu et al., 2024).
• Adaptive and active attackers: Most frameworks currently assume passive eavesdroppers; adaption to machine-learning–empowered, side-channel, or query-based attacks is an emergent challenge (Tang et al., 29 Mar 2025).
• Extension to multi-modal, multi-user, and federated settings: Scaling frameworks to distributed, cross-modal, or federated environments, possibly with dynamic coalition formation or incentive alignment, is an active research direction (Yu et al., 5 May 2026, Liu et al., 2 Mar 2026).

Research is ongoing into synergizing semantic code design with multi-agent coordination, extending coverless and token-controlled methods for diverse modalities, and developing RL- and contract-theoretic optimization schemes that account for cognitive asymmetry and adversarial learning (Xu et al., 23 May 2025, Meng et al., 23 Jan 2026, Liu et al., 2 Mar 2026, Wu et al., 9 Apr 2026, Liu et al., 19 Jan 2025).