Deep neural networks (DNNs) are extremely vulnerable to adversarial attacks, where small changes to the input, unnoticeable to humans, can drastically alter the model’s predictions. One reason for their poor performance is that they process as much information as possible and make decisions based on the input received. Human vision performs differently. It tries be as efficient with information as possible while using predictions to fill in missing information based on existing knowledge. The brain may even suppress sensory input to save energy if it is familiar (i.e., predictable). As the knowledge follows the average distribution of past experiences, it tends to be correct and unbiased. Therefore, perturbations applied solely to the input are unlikely to affect brain function without being noticed. We hypothesize that predictive sensing and perception are key to the robustness of human sensory systems. One of the features that is unique in human vision but lacking in the artificial vision systems is the foveal-peripheral vision sampling with saccadic eye movements. This is an attention mechanism that sparse samples unimportant (i.e., predictable) regions while dense samples regions of interests (i.e., unpredictable) or with saliency features. To test our hypothesis, we developed a foveal-peripheral sampling front-end with saccades and integrate it with the traditional DNN using predictive reconstruction. Some existing works have also studied the foveal-peripheral visual mechanisms, however, they do not have a saccadic system to actively collect useful information and they do not fill-in the missing (predictable) information using prediction. In this work, we first demonstrate that foveal-peripheral sampling with saccades exhibits strong robustness to adversarial attacks when using random saccades. Next, we show that the framework with trained saccades could further enhance the model’s robustness, which confirms the importance of using a trained saccadic system with foveal-peripheral vision.