Manipulating Digital Twin Networks by Poisoning Semantic Knowledge Base

With the development of digital twins (DT) and the metaverse, network virtualization is leading to an increase in NPCs. Deep learning-based semantic communication systems with the knowledge base (KB) are key to enhancing their interaction and communication. This development is expected to optimize routing automation and network service quality, becoming a key trend in the evolution of future communication networks. However, the security of the KB in semantic communication remains a critical challenge. This paper explores the vulnerabilities of semantic communication systems to backdoor attacks, focusing on their impact on the KB. We propose a method for injecting malicious data via backdoor triggers in the KB, which affect message recovery during communication. Experimental results show that semantic distortion degree is around 0.4. Under Additive White Gaussian Noise (AWGN) channel conditions, the attack success rate (ASR) reaches 99% at the 10 dB Signal-To-Noise Ratio (SNR), the performance of the proposed method outperforms the previous approach within the 0–10 dB SNR range, and 90% and 98% in Rayleigh and Rician channels, respectively, while the proposed approach will not affect the normal communication without triggers. These findings highlight the need for robust security mechanisms to protect the KB in semantic communication systems, particularly in DT network.