<p>Masked image modeling (MIM) has shown remarkable prowess in self-supervised representation learning. While reconstructing corrupted images has emerged as an alternative pretext task, existing methods often rely on generic corruptions generated by auxiliary models, incurring significant computational costs and lacking targeted difficulty relevant to the pre-training objective. To address these limitations, we propose Adversarial Examples Meet Masked Image Modeling (AEMIM), a novel framework that incorporates adversarial examples (AEs) as challenging reconstruction inputs. AEs are generated online using the model being trained, eliminating the need for extra generators and enhancing efficiency. Specifically, AEMIM introduces an auxiliary pretext task focused on reconstructing AEs corresponding to the original masked images, complementing the primary MIM task. Crucially, we devise an innovative adversarial attack strategy tailored for MIM, utilizing the feature-space distance between clean and adversarial encoder representations as the adversarial loss. This directly targets encoder robustness and avoids pitfalls associated with using reconstruction loss for AE generation. Adapters are employed to facilitate stable co-training with clean and adversarial data streams. Our method acts as a versatile plug-in, compatible with various MIM strategies and model architectures. Comprehensive experiments demonstrate that AEMIM significantly enhances the generalization and robustness of pre-trained models. Notably, AEMIM surpasses strong baselines like MAE on ImageNet classification, its IID and OOD variants, and downstream tasks including object detection, instance segmentation, and semantic segmentation. A fast version, Fast AEMIM, is also presented, offering a compelling trade-off between performance and efficiency.</p>