Abstract
Deep learning (DL)-driven receivers have revolutionized wireless communications but remain vulnerable to over-the-air (OTA) adversarial attacks. To address this challenge, this paper introduces an attack-agnostic, test-time adaptive frame-work for robust wireless reception. The core of this framework leverages the rich data distribution learned by a pre-trained diffusion model to purify adversarially perturbed inputs. We achieve this by projecting inputs back onto the original data manifold through a likelihood maximization (LM) process guided by denoising gradients. Concurrently, this purification process is integrated with self-ensemble inference to improve decision-making reliability. Performance evaluations demonstrate that, compared to baseline methods, the proposed framework achieves an order-of-magnitude reduction in bit error rate (BER) across a spectrum of attacks. This highlights its strong generalization, offering a promising defense solution for next-generation secure intelligent communication systems.