Wireless networks using AI for resource allocation fail when deliberately jammed.

Researchers found that adversarial jamming attacks can trick AI systems managing 5G network slices into violating service agreements, and that the AI takes significant time to recover from these attacks. In practice, this means a small budget for targeted interference can degrade network performance for specific users or services, and the system won't bounce back immediately.

5G networks are increasingly automated — they use machine learning to decide which users get which radio resources in real time. This paper shows the system has a specific vulnerability: an adversary doesn't need to jam everything, just enough to poison the AI's training data and force it to make bad allocation decisions. The vulnerability exists because the AI learns by trial and error, and if those early trials are under attack, it learns the wrong thing. Recovery is slow because the AI has to unlearn bad patterns before it can adapt to normal conditions again.