MaskAug: Architecture-agnostic corruption-aware training for robust network intrusion detection

Abstract

Deployed network intrusion detection systems frequently encounter corrupted or missing features due to sensor failures, network disruptions, and resource constraints: conditions that laboratory benchmarks rarely capture. Existing approaches typically assume complete feature availability and couple robustness improvements to specific architectures. We propose mask augmentation ( MaskAug ), a data-level corruption-aware training strategy that simulates deployment-time feature degradation through three complementary operators: random masking, block masking, and proportional noise injection. Because MaskAug operates on the input before the classifier, it integrates with any differentiable neural backbone as a drop-in augmentation, without modifying the inference pipeline. We evaluate MaskAug on leakage-audited benchmarks rebuilt directly from the official UNSW-NB15 and CICIDS2017 corpora, using 10 independent train/evaluation splits with three model seeds each. Across multilayer perceptron (MLP), one-dimensional convolutional neural network, and Feature Tokenizer + Transformer, MaskAug improves corruption-averaged Macro-F1 by + 0.043 / + 0.060 / + 0.072 on UNSW-NB15 and + 0.087 / + 0.150 / + 0.057 on CICIDS2017 over clean-trained baselines, and the gains persist against a budget-matched control trained on replicated clean samples (equalizing optimizer updates): all 12 pre-specified split-level one-sided Wilcoxon comparisons remain significant after Holm correction (adjusted p = 0.012 ), with effect sizes reported via cluster-bootstrap confidence intervals. A lexical term frequency-inverse document frequency baseline remains competitive on UNSW-NB15 yet collapses under masking on CICIDS2017 (Macro-F1 0.33 vs. 0.92 for MLP+MaskAug). Corruption-aware training thus delivers architecture-agnostic, cross-dataset-consistent robustness at zero inference-time cost.