Generative AI-driven Cross-layer Covert Communication: Fundamentals, Framework and Case Study

TL;DR

The paper addresses end-to-end covert communication security in military networks by proposing a GenAI-driven cross-layer framework that combines physical, network, and application-layer techniques. It introduces a channel-space–based AI engine to construct end-to-end covert paths, aided by Generative Diffusion Models (GDM) and reinforcement learning to model wardens and optimize channel selection. A diffusion-empowered reinforcement learning approach (DSAC) is demonstrated in a CEIoT case study, showing superior channel-quality optimization $H(\sigma, P_D) = V \sigma (1 - P_D)$ under reachability and covert constraints. The work contributes a comprehensive framework, methodological advances in GenAI-based cross-layer concealment, and a practical diffusion-enabled case study, with implications for robust, low-detectability communications in complex multi-layer networks.

Abstract

Ensuring end-to-end cross-layer communication security in military networks by selecting covert schemes between nodes is a key solution for military communication security. With the development of communication technology, covert communication has expanded from the physical layer to the network and application layers, utilizing methods such as artificial noise, private networks, and semantic coding to transmit secret messages. However, as adversaries continuously eavesdrop on specific communication channels, the accumulation of sufficient data may reveal underlying patterns that influence concealment, and establishing a cross-layer covert communication mechanism emerges as an effective strategy to mitigate these regulatory challenges. In this article, we first survey the communication security solution based on covert communication, specifically targeting three typical scenarios: device-to-device, private network communication, and public network communication, and analyze their application scopes. Furthermore, we propose an end-to-end cross-layer covert communication scheme driven by Generative Artificial Intelligence (GenAI), highlighting challenges and their solutions. Additionally, a case study is conducted using diffusion reinforcement learning to sovle cloud edge internet of things cross-layer secure communication.

Figures (5)

• Figure 1: The characteristics of covert communication schemes in different layers.
• Figure 2: The framework of GenAI-driven cross-layer covert communication.
• Figure 3: The illustration of cross-layer covert communication scheme in CEIoT, the influence of warden is detection ability $d$ and discount factor $\alpha$, which physically means warden's ability to detect data in the channel. Cross-layer multi-channel supervision dilutes its detection ability for each channel, so detection ability needs to be multiplied by $\alpha$.
• Figure 4: Diffusion empowered reinforcement learning model
• Figure 5: Numerical results of our diffusion-empowered SAC and conventional SAC