<p>We propose the Adaptive Masked Autoencoder (AMA), a reconstruction-based anomaly detector for structured tabular data. AMA extends masked autoencoders by replacing static random masks with adaptive masks driven by two signals: global feature importance and feature-level reconstruction difficulty. Global feature importances are obtained from a supervised model trained on labeled normal and anomalous samples, while reconstruction errors are updated online during autoencoder training. The masking ratio is gradually increased over epochs, forming a simple curriculum that first exposes the model to lightly corrupted inputs and then progressively emphasizes more challenging features. We evaluate AMA on five benchmark datasets, including NSL-KDD, UNSW-NB15, TON-IoT, CICIDS2017, and a credit card fraud dataset, using validation-based hyperparameter selection and reporting mean±std over five runs. Across these benchmarks, AMA achieves competitive results and, in several network-intrusion settings, improved AUC-ROC and/or F1-scores relative to the vanilla autoencoder and the fixed-mask masked autoencoder, with the largest gains on more challenging protocol subsets and low-signal attacks. These results suggest that adaptive, feedback-driven masking is a promising direction for strengthening autoencoder-based anomaly detection in high-dimensional tabular domains.</p>