Spatially Distributed Multi-Target Backdoor Attacks Against Automatic Modulation Recognition

Automatic modulation recognition based on deep learning has become a crucial technology for wireless communication systems. However, the processes involved in open data collection and the outsourcing of training by developers may introduce security vulnerabilities, making these systems susceptible to significant threats from backdoor attacks. Previous research has primarily concentrated on enabling attackers to manipulate inputs to target a single modulation type, which has limited relevance to the complexities of real-world wireless communication networks. This paper introduces a novel method for spatially distributed multi-target backdoor attacks, which embeds triggers at various spatial locations within the modulation signal to simultaneously target multiple modulation types. In contrast to traditional single-target attacks, our approach enhances both the stealth and robustness of the attack by distributing poisoned samples across several target classes. Experimental results indicate that our method achieves a higher average attack success rate compared to existing techniques, while also preserving accurate classification performance for benign samples. Additionally, it effectively circumvents current backdoor detection mechanisms, underscoring its robustness. This study highlights a new category of security threats confronting automatic modulation recognition in wireless communication systems and serves as a significant reference for the advancement of future defense technologies.