Missing values are common in time-series anomaly detection, which can degrade performance when adversaries maliciously exclude critical features or time steps. Previous methods, like interpolation and imputation, cannot guarantee robustness against such data exclusion attacks. To address this challenge, we introduce a novel adversarial training framework that leverages strategic masking to enhance the robustness of anomaly detectors. Our training method can be formulated as a zero-sum game, where the anomaly detector aims to minimize the worst-case loss against an adversary that seeks to maximize it with well-crafted adversarial samples. The minimization process of the anomaly detector is essentially a deep-learning imputation step on the attacked features, which is more challenging than traditional imputation methods due to the adversarial nature of the attack. Our future work will focus on generalizing the proposed adversarial training framework under the certified robustness theory, which guarantees the model scoring consistency under data exclusion attacks within a bounded region.