VocoX: Adversarial Attack Against AI-based Voice Control Services in Semantic Communications

Semantic communications are expected to be the dominant technology in next-generation networks. This technology provides a receiver with semantic information obtained from the source, allowing for transmission throughput to exceed Shannon’s theoretical capacity limit. This capability is crucial in bandwidth-intensive systems, such as digital twins and metaverse. Nevertheless, the susceptibility of semantic communications to adversarial perturbations is further amplified by the vulnerability of neural networks and the unrestricted nature of wireless channels. This work presents a novel adversarial white-box model that exploits the sensitivity of wireless channels to manipulate the meaning of the original semantic information through artificial noise generation and injecting it into the voice waveform. The experimental results show that the attack can cause a degradation of up to 80% in the speech quality of voice control services in semantic communications. Although there is no perfect defense approach, adversarial training can be a cost-effective solution to improve the resilience of intelligent voice control services in semantic communications against adversarial attacks.