Abstract
This paper investigates backdoor attacks in image-oriented semantic communications. The threat of backdoor attacks on symbol reconstruction in semantic communication (SemCom) systems has received limited attention. Previous research on backdoor attacks targeting SemCom symbol reconstruction primarily focuses on input-level triggers, which are impractical in scenarios with strict input constraints. In this paper, we propose a novel channel-triggered backdoor attack (CT-BA) framework that exploits inherent wireless channel characteristics as activation triggers. Our key innovation involves utilizing fundamental channel statistics parameters, specifically channel gain with different fading distributions or channel noise with different power, as potential triggers. This approach enhances stealth by eliminating explicit input manipulation, provides flexibility through trigger selection from diverse channel conditions, and enables automatic activation via natural channel variations without adversary intervention. We extensively evaluate CT-BA across four joint source-channel coding (JSCC) communication system architectures and three benchmark datasets. Simulation results demonstrate that our attack achieves near-perfect attack success rate (ASR) while maintaining effective stealth. Finally, we discuss potential defense mechanisms against such attacks.
Introduction
As the telecommunications industry moves toward 6G, Semantic Communication (SemCom) is emerging as a transformative technology. Unlike traditional systems that focus on bit-perfect transmission, SemCom leverages Deep Learning (DL) to transmit the "meaning" of data, achieving significant compression and efficiency. However, the reliance on DL models introduces new security risks, particularly backdoor attacks.
Existing backdoor research in SemCom has largely focused on input-level triggers, such as adding specific pixel patches to images. These methods are often impractical as they require access to the source data and can be detected through input sanitization. This paper addresses these limitations by proposing the Channel-Triggered Backdoor Attack (CT-BA), which uses the physical wireless environment itself as the trigger switch.
Methodology
The core innovation of CT-BA is the use of wireless channel characteristics—such as Signal-to-Noise Ratio (SNR) or fading distributions—as an "activation switch." Because modern SemCom systems utilize end-to-end (E2E) training with differentiable channel layers, the model can be trained to associate specific physical conditions with a malicious response.
The authors designed two primary trigger strategies:
- n-trigger (Noise-based): The backdoor activates at a specific noise power level or SNR.
- H-trigger (Gain-based): The backdoor activates based on specific channel gain statistics, such as a particular fading distribution (e.g., Rayleigh fading).
During the training phase, the attacker "poisons" a small portion of the training data (as low as 1%). The model learns to perform standard reconstruction under normal conditions but switches to a pre-defined target image whenever the transmitted signal encounters the trigger channel conditions.
Experimental Evaluation and Results
The researchers tested CT-BA using three benchmark datasets: MNIST, CIFAR-10, and ImageNet. The attack was evaluated across four Joint Source-Channel Coding (JSCC) architectures, including both CNN-based systems (such as BDJSCC and ADJSCC) and modern Vision Transformer (ViT)-based systems.
The proposed attack demonstrated exceptional performance and stealth:
- Attack Success Rate (ASR): Reached 100% in nearly all scenarios, indicating the target image was perfectly reconstructed upon trigger activation.
- Stealth: The poisoned model's performance on clean channels was virtually indistinguishable from a non-poisoned model, with PSNR differences often below 0.5%.
- Robustness: The attack remained effective across various bandwidth ratios and low poisoning rates.
To counter this threat, the authors proposed a "Noise Titration" defense. By injecting controlled noise into the decoder, defenders can observe if the output suddenly jumps to a suspicious target image, effectively exposing the hidden backdoor.
Conclusion
CT-BA represents a highly stealthy and dangerous new class of physical-layer attacks on semantic communications. By moving the trigger from the digital to the physical domain, attackers can exploit natural channel variations to activate malicious behaviors without direct intervention. Future research will focus on more complex channel dynamics and extending these findings to text, speech, and video-based SemCom systems.
Summary updated