On the Vulnerability of Deep Automatic Modulation Classifiers to Explainable Backdoor Threats

Backdoor attacks on wireless AI systems are a serious threat: Deep learning models used for automatic modulation classification (AMC)—a key technology in modern wireless communications—can be successfully compromised through backdoor attacks that work across a range of signal quality conditions and require only small amounts of poisoned training data.

Explainable AI can be weaponized to make attacks more effective: The researchers demonstrated that adversaries can use explainable AI (XAI) techniques to identify the most vulnerable parts of wireless signals, allowing them to design more efficient backdoor triggers that are harder to detect.

Physical-layer attacks represent a new frontier in AI security: Unlike most prior research on backdoor attacks in computer vision, this work shows that adversaries can inject malicious triggers directly into wireless signals before transmission, creating a novel attack vector that wireless communication systems must defend against.

Current AI models lack adequate defenses for this threat: The study found that multiple deep learning-based AMC models were vulnerable to these backdoor attacks, suggesting that existing security measures may be insufficient for AI systems deployed in critical wireless infrastructure.

Small amounts of compromised data can cause widespread failures: The attacks achieved high success rates with only a small poisoning ratio, meaning adversaries don't need to compromise large portions of training data to successfully manipulate these AI systems.