Image Steganography For Securing Intellicise Wireless Networks: "Invisible Encryption" Against Eavesdroppers

TL;DR

The paper addresses security and privacy challenges in SemCom for intellicise wireless networks by proposing image steganography as an invisible encryption approach. It surveys encryption schemes, surveys and categorizes image steganography paradigms, and analyzes CNN-, GAN-, and INN-based JSCC models for secure SemCom, complemented by six training strategies. A case study demonstrates a conditional diffusion-based coverless Steganography SemCom scheme, showing key-based recovery and discussing its advantages and limitations. The work outlines future directions including theoretical capacity analysis, multimodal steganography, GAI-based steganalysis, and hybrid encryption-steganography approaches to advance secure SemCom. Together, these contributions offer a practical, end-to-end framework for concealing semantic data and defending against intelligent eavesdroppers in SemCom-enabled networks.

Abstract

As one of the most promising technologies for intellicise (intelligent and consice) wireless networks, Semantic Communication (SemCom) significantly improves communication efficiency by extracting, transmitting, and recovering semantic information, while reducing transmission delay. However, an integration of communication and artificial intelligence (AI) also exposes SemCom to security and privacy threats posed by intelligent eavesdroppers. To address this challenge, image steganography in SemCom embeds secret semantic features within cover semantic features, allowing intelligent eavesdroppers to decode only the cover image. This technique offers a form of "invisible encryption" for SemCom. Motivated by these advancements, this paper conducts a comprehensive exploration of integrating image steganography into SemCom. Firstly, we review existing encryption techniques in SemCom and assess the potential of image steganography in enhancing its security. Secondly, we delve into various image steganographic paradigms designed to secure SemCom, encompassing three categories of joint source-channel coding (JSCC) models tailored for image steganography SemCom, along with multiple training strategies. Thirdly, we present a case study to illustrate the effectiveness of coverless steganography SemCom. Finally, we propose future research directions for image steganography SemCom.

Figures (4)

• Figure 1: Illustration of image steganography SemCom systems, where (a) presents the basic architecture of the image steganography system, (b) CNN-based image steganography JSCC model has lowest training time and model size but worst secret semantic capacity and transmission quality, (c) GAN-based image steganography JSCC model has the best anti-eavesdropping ability through adversarial training but the most parameters, and (d) INN-based image steganography JSCC model has the highest secret semantic capacity and transmission quality but longest training time.
• Figure 2: Illustration of six image semantic steganography strategies. (a) Cost-based strategy leverages reinforcement learning to optimize embedding decisions with minimal cost, making it ideal for scenarios with limited steganographic capacity. (b) Adversarial strategy integrates adversarial examples during training to enhance resilience against detection by semantic steganalyzers. (c) Self-learning strategy utilizes end-to-end DL models to enable autonomous semantic steganography, eliminating the need for manual preprocessing or post-processing steps. (d) High-frequency region strategy embeds secret semantic information into high-frequency domains, leveraging regions with rich cover semantic features to evade semantic steganalysis. (e) Pre-trained model strategy avoids time-consuming model retraining, thus enhancing steganographic efficiency. (f) Coverless strategy generates stego images directly through GAI models, bypassing the requirement for cover images.
• Figure 3: Illustration of the proposed coverless steganography SemCom scheme, where the conditional diffusion-based coverless steganography module and JSCC-based semantic codec are trained separately, and keys are generated based on semantic features. The public keys are open access to everyone, including intelligent eavesdroppers, and they are semantic information related to steganographic images but not to secret images.
• Figure 4: Simulation results of the proposed conditional diffusion-based coverless steganography SemCom scheme. We select the Stable Diffusion version 1.5 as the conditional diffusion model and employ the deterministic DDIM as the sampling algorithm. Both the forward and reverse processes are configured to include 50 steps. The DeepJSCC architecture is employed as the trained semantic encoder and decoder.