SecureSemComs: Dynamic Adversarial Defense Framework for Semantic Communications
Shuwen Zhang, Siye Wang, Jiaxing Li · 2025
Semantic communication is a current research hotspot emphasizing the transmission of content-related semantics. However, this focus increases vulnerability to attacks that exploit high-dimensional features, underscoring the urgent need for robust defense mechanisms against evolving threats. Traditional defense strategies are often computationally intensive, require continuous updates, and perform inadequately against sophisticated adversarial tactics in noisy environments. In this paper, we propose a Dynamic Adversarial Defense Framework (DADF), a model-agnostic, plug-and-play solution for seamless integration with semantic communication systems, requiring no modifications for deployment. The Adversarial Pattern Learning (APL) module aligns adversarial and genuine audio samples within a unified feature space, facilitating noise pattern extraction and significantly reducing interference. Subsequently, a Dynamic Noise Filter (DNF) employs adaptive mechanisms to eliminate harmful noise based on noise patterns extracted from the APL. Testing across five public benchmarks demonstrates DADF's superior performance, with peak achievements including a 9.31-fold reduction in mean squared error (MSE) and a 269.78% increase in signal-to-distortion ratio (SDR), significantly outperforming existing solutions. This breakthrough establishes a new bench-mark for resilience and adaptability in semantic communications, enhancing security against evolving adversarial attacks. The code is available at https://github.com/sweet0516IDADF.