Agentic AI-driven SemSteCom

• Agentic AI-driven SemSteCom is a framework that integrates AI-based semantic extraction, token-controlled reference image generation, and coverless steganography to secure wireless communications.
• It leverages a modular architecture comprising semantic extraction, EDICT-based steganography, a semantic JSCC codec, and optional enhancement to optimize transmission and recovery.
• Benchmark results demonstrate significant improvements in PSNR, SSIM, and MSE, confirming robust semantic security and resilience against eavesdropping.

Agentic AI-driven SemSteCom (AgentSemSteCom) denotes a paradigm of secure and efficient semantic steganographic communication for future intellicise wireless networks, in which agentic artificial intelligence orchestrates semantic extraction, generative steganography, semantic channel coding, and task-based enhancement modules. AgentSemSteCom systematically eliminates the requirements for both cover images and private semantic keys, thereby maximizing steganographic capacity and transmission security. The technical foundation integrates agent-controlled semantic feature extraction, digital token–regulated reference image synthesis, mathematically exact invertible diffusion sampling, and advanced semantics-aware joint source-channel coding under agentic feedback and optimization (Meng et al., 23 Jan 2026, Yu et al., 1 Oct 2025, Gao et al., 29 Dec 2025).

1. System Architecture and Workflow

AgentSemSteCom architecture comprises five principal modules: semantic extraction, digital-token-controlled reference image generation, coverless steganography, semantic codec, and optional task-oriented enhancement. Decision-making and orchestration across modules are governed by agentic AI entities such as LLMs, large vision models (LVMs), and RL-based control agents. The workflow proceeds as follows (Meng et al., 23 Jan 2026):

1. Semantic Extraction: Agentic AI extracts a public semantic key $K_{pub}$ (via LLM) and implicit semantic features $x_{feat}$ (via domain-specific deep networks) from the secret image $x_s$ and requirements $R$.
2. Digital Token Reference Generation: A user-provided token $k$ is hashed to seed a latent noise $z_D$. A ControlNet-augmented diffusion model, conditioned on $K_{pub}$ and $x_{feat}$, generates a reference image $x_{ref}$ via DDIM sampling and VAE decoding.
3. Coverless Steganography: The secret $x_s$ is VAE-encoded to latent $z_s$. EDICT-based invertible forward-noising is applied, token-guided perturbation via binary mask $M(s, \eta)$ is performed, then conditional reverse EDICT denoising (with decoupled cross-attention on $K_{pub}$ and $x_{ref}$) reconstructs the stego-image $x_{stego}$.
4. Semantic Codec Transmission: $x_{stego}$ is encoded by a semantic JSCC codec, transmitted over AWGN channels, and decoded at the receiver by $D_{sem}$.
5. Optional Enhancement: Agentic AI selects enhancement operator(s) (e.g., face denoiser, super-resolution) to fulfill $R$ on recovered secret $x_s^*$.

The diagram below outlines the principal order:

Step	Module	Output
1	Semantic Extraction	$K_{pub}$, $x_{feat}$
2	Token Reference Gen.	$x_{ref}$
3	Steganography	$x_{stego}$
4	Semantic Codec	Transmitted $x_{stego}$
5	Enhancement (optional)	$x_s^*$

2. Mathematical Models

Semantic Feature Extraction

• Public Semantic Key ($K_{pub}$): Autoregressive LLM yields $q_{\ell}$ at each layer via:

$$q'_{\ell} = CSA(LN(q_{\ell-1})) + q_{\ell-1} \ q_{\ell} = MLP(LN(q'_{\ell})) + q'_{\ell}$$

• Token Sampling: For $i=1,\ldots,T$,

$$\hat{K}_i = \text{Softmax}(W_{pub} \cdot LN(q_L^i)), \quad K_i \sim \hat{K}_i$$

Form $K_{pub} = (K_1, \ldots, K_T)$.

• Implicit Features ($x_{feat}$): $x_{feat}=\mathcal{G}(x_s, R)$, e.g., OneFormer segmentation for general, OpenPose for faces, $x_{feat}=\varnothing$ for style-only.

Reference Image Generation via Diffusion

• Token-controlled latent $z_D = \text{Randn}(s)$, $s = \mathcal{H}(k)$ (hashed user token).
• Diffusion sampling: ControlNet-augmented U-Net backbone guided by $K_{pub}$ and $x_{feat}$:
  • DDIM iterative update:

  $$z_{t-1} = \sqrt{\alpha_{t-1}} \left( \frac{z_t - \sqrt{1-\alpha_t} \hat{\epsilon}_\theta}{\sqrt{\alpha_t}} \right) + \sqrt{1-\alpha_{t-1}} \hat{\epsilon}_\theta$$ - Final $x_{ref} = D_{VAE}(z_0)$.

• Training Loss:

$$\mathcal{L} = \mathbb{E}_{z_0, t, x_{feat}, K_{pub}, \epsilon}[ \| \epsilon - \epsilon_\theta(z_t, t, x_{feat}, K_{pub}) \|_2^2 ]$$

EDICT-based Steganography

• VAE encoding: $z_s = E_{VAE}(x_s)$

• Forward EDICT invertible noising: $(z_0,u_0) = (z_s,z_s)$, affine coupling and mixing as per eqs. 14a–14d (Meng et al., 23 Jan 2026)

• Token-guided perturbation: $z'_T = z_T \odot (1-M) - z_T \odot M$

• Decoupled cross-attention (reverse):

  • $Z_{text}$ and $Z_{img}$ calculated with $K_{pub}$ and $x_{ref}$ embeddings at each U-Net layer, summed to give $Z_{con}$.
• EDICT-DDIM update:

$$z_{t-1} = \alpha_t (p z_t + (1-p) u_t) + \beta_t \tilde{\epsilon}_{t,c}$$

with $\tilde{\epsilon}_{t,c}$ compounded by $K_{pub}$, $x_{ref}$.

JSCC Semantic Codec

• Encoder: $S = E_{sem}(x_{stego})$
• Channel Model: $\hat{S} = h \cdot S + n$, $n \sim \mathcal{CN}(0,\sigma_w^2)$
• Decoder: $\hat{x}_{stego} = D_{sem}(\hat{S})$
• Objective: Analog variant minimizes $$\mathcal{L}_{sem} = \mathbb{E}[d(x_{stego}, \hat{x}_{stego}) + \lambda \|S\|^2]$$

3. Integrated Agentic AI Control and Semantic Communication

Agentic AI in SemSteCom fulfills dual objectives: strategic module selection (semantic extraction, enhancement) and secure protocol orchestration (token-key management, attack resistance). Control agents include:

• LLM-Driven Key Extraction: Ensures $K_{pub}$ derives only public, non-reversible semantics.
• Digital Token Management: User token $k$ deterministically seeds reference generation and stego encoding; wrong token collapses EDICT invertibility.
• Feature Selection Strategy: Contextual networks such as segmentation or pose models augment extraction based on task $R$.
• JSCC Codec Orchestration: Embedded agents optimize encoding parameters for semantics fidelity and channel conditions (Yu et al., 1 Oct 2025, Gao et al., 29 Dec 2025).

A plausible implication is that future extensions may permit active negotiation of semantic protection levels and adaptive stego capacities in agentic networks.

4. Security Analysis and Steganographic Capacity

AgentSemSteCom addresses semantic eavesdropping threats by:

• Elimination of Cover Images: Stego image $x_{stego}$ bears no overt payload markers, and external cover is unnecessary.
• Token-Guided Invertibility: Recovery of $x_s$ requires precise token $k$ and semantic features; an incorrect token results in non-recoverable outputs (ΔPSNR up to 6.5 dB between legitimate receiver and agentic-AI eavesdropper at SNR=10 dB for facial class) (Meng et al., 23 Jan 2026).
• EDICT Mathematical Exactness: At each step, affine updates in EDICT chains are bijective, ensuring forward noising is invertible only with correct agentic keys.
• Steganographic Capacity: Full-resolution semantics are embeddable; capacity is quantifiable as $C_{up} \approx H \cdot W \cdot \text{channels}$.

Privacy is further enhanced by adding differential privacy noise to semantic embeddings $s(x)$, subject to distortion $\Delta_{sem} \leq \epsilon_{priv}$ constraints (Yu et al., 1 Oct 2025).

5. Quantitative Performance and Benchmarks

On the UniStega dataset (100 images, three content classes):

Method	PSNR (dB)	SSIM	MSE ($\times 10^{-2}$)	LPIPS
AgentSemSteCom	21.59	0.5643	0.799	0.4030
SemSteDiff	18.86	0.5183	1.21	0.4046

AgentSemSteCom demonstrates +14.29% PSNR, +8.88% SSIM, −43.75% MSE, and −5.1% LPIPS improvements over SemSteDiff (Meng et al., 23 Jan 2026). For eavesdropper scenarios (wrong token), PSNR drops by ~6.5 dB, confirming significant resilience to semantic interception. Transmission over AWGN channels at SNR ∈ {5,10,15,20} dB further validates robust reconstruction.

The AKB-JSCC framework (source/channel KB) is shown to stably outperform NTSCC and traditional digital codecs in MS-COCO image task, yielding higher PSNR under both bandwidth and channel SNR constraints (Gao et al., 29 Dec 2025).

6. Applications, Deployment, and Future Directions

AgentSemSteCom is applicable to autonomous drone swarms, multi-vehicle perception, and multi-robot rescue in 6G and intellicise networks:

• Autonomous Drones: Fine-tuning of semantic encoders via GAN synthesis, lightweight pruned/quantized models, and hierarchical RL for spectrum allocation maximize coverage/QoE (Yu et al., 1 Oct 2025).
• Secure Multi-Agent Coordination: Coverless steganography allows secret intent or commands to be embedded and exchanged securely, with optional post-recovery enhancement based on downstream tasks.
• Challenges: Heterogeneity of agents and rapid topology changes are mitigated by backward-compatible model distillation and warm-started model banks.
• Standardization: Interoperability requires unified semantics-API, signaling protocols, and agent KPIs.

Extension to broader semantic communication scenarios includes continual learning for evolving knowledge bases, modular testbeds for reproducibility, and integrating federated privacy protocols (Gao et al., 29 Dec 2025).

7. Summary and Outlook

AgentSemSteCom manifests the convergence of agentic AI with semantic communication, generating a knowledge-augmented, security-hardened, and resource-optimal framework for task-oriented wireless intelligence. The principal advancements—coverless steganography via token-controlled invertible diffusion, agentic orchestration of feature extraction and codec adaptation, and exact semantic recovery—set a new benchmark for confidential, high-capacity, low-latency information exchange. Emerging research focuses on robust semantic intent signaling, modular agent deployment, and dynamic evolution of semantic-code and knowledge-base structures (Meng et al., 23 Jan 2026, Yu et al., 1 Oct 2025, Gao et al., 29 Dec 2025).