The rapid proliferation of wireless Internet of Things devices demands reliable and secure connectivity, which can be achieved by Wireless Link Estimation (WLE) models, particularly those based on deep learning. However, these models are vulnerable to dropout attacks. This attack manipulates dropout layers to degrade model accuracy, leading to inappropriate or malicious network selections. In this paper, we introduce a dual-layer defense mechanism with real-time anomaly detection within the dropout layer. This mechanism monitors gradient variations to detect dropout attacks and rapidly reverts the compromised layers to a secure state, thereby ensuring stable training. Extensive experimental evaluation across various datasets and neural architectures highlights the severity of dropout attacks, which drop model accuracy to nearly 20%, and demonstrates the effectiveness of our solution, which restores the accuracy of WLE models to near pre-attack levels.