The researchers argue that intrusion detection systems have long suffered from a hidden flaw in their data pipelines. Most published studies apply a single uniform scaling method, usually min-max normalization, to all features at once. That approach works reasonably well when every feature has a similar statistical character, but it breaks down badly when a dataset mixes features with wildly different distributions, as modern security datasets increasingly do. Because preprocessing is typically fixed by convention rather than evaluated, the authors describe it as a hidden hyperparameter, one whose cost is large enough in their experiments to change conclusions about which models are suitable.

The scale of the mismatch is striking. In the ToN-IoT benchmark, which combines network flow statistics with host-level and IoT device telemetry, feature magnitudes span nine orders of magnitude, from binary flags taking values of zero or one to byte counts exceeding four billion. When min-max normalization squeezes both into the same range, the enormous feature dominates the learning signal and the binary feature is effectively crushed into noise. The study shows that this crushing has real consequences: under the baseline, the model produced false alarms on 67.1 percent of normal traffic on ToN-IoT, a rate that fell to 5.6 percent after semantic conditioning was applied, a twelvefold reduction.

The problem is compounded in datasets with protocol-split structures. Edge-IIoTset aggregates features from eight distinct protocols, including TCP, UDP, ICMP, DNS, HTTP, ARP, MQTT, and Modbus/TCP. For any individual record, features belonging to non-active protocols are uniformly zero, producing a sparse matrix in which uniform min-max scaling assigns near-identical normalized values to the majority of features regardless of their semantic content. After dropping columns with more than 95 percent zeros, only 24 features remained for 15 classes, and the baseline model managed just 38.7 percent accuracy on that dataset.

MAFT works in two stages. First, it classifies each feature into one of six semantic groups: temporal, volumetric, behavioral, identity, protocol-specific, or statistical. The classification uses a hybrid approach, combining deterministic rule-based pattern matching on feature names, such as patterns for duration, byte counts, or protocol flags, with statistical clustering on a six-dimensional distributional profile covering mean, standard deviation, skewness, kurtosis, unique value ratio, and zero fraction. When the two methods disagree, the rule-based assignment wins, because feature names in security datasets, which follow conventions from tools like CICFlowMeter, Argus, and Zeek, tend to encode genuine domain meaning that value distributions alone cannot recover.

Second, each semantic group receives a scaling method matched to its statistical character. Temporal features, which tend to follow approximately Gaussian distributions, receive standard z-score scaling. Volumetric features, which are typically heavy-tailed and dominated by extreme outliers, receive robust scaling based on the median and interquartile range, followed by clipping of extreme values to a bounded range that prevents gradient explosion during training. Behavioral, identity, protocol-specific, and statistical features, which are bounded or sparse, receive min-max scaling that preserves their original zero-to-one or ordinal structure. A fallback cascade handles sparse features whose interquartile range is exactly zero, a situation that arises frequently in protocol-split datasets.

The evaluation covered six widely used public benchmark datasets spanning enterprise network traffic, IoT environments, IoT botnet detection, and industrial control systems, with feature counts ranging from 24 to 115 and class counts from 6 to 15. The team ran a four-stage ablation study comprising 24 centralized experiments in which the model architecture was held constant while the data preparation method varied, allowing the effect of semantic conditioning to be isolated from any effect of model design. All normalization methods were fitted exclusively on training splits and applied to validation and test data without refitting, preventing distributional information from the test data from influencing the reported results.

The results split cleanly along a line the authors call feature heterogeneity. Datasets that mixed features from many sources and scales showed the largest gains, with single-seed multiclass accuracy improvements of 16.1 and 22.9 percentage points on ToN-IoT and Edge-IIoTset respectively, the two most heterogeneous datasets. Moderately heterogeneous enterprise datasets, UNSW-NB15 and CICIDS2017, gained 3.8 and 1.7 percentage points. Datasets whose features came from a single source with similar statistical character, such as the HAI industrial control dataset with its 61 continuous process variables and the N-BaIoT botnet dataset with its 115 uniformly computed stream statistics, showed essentially no change, because uniform scaling was already adequate.

Beyond raw accuracy, the study reports several secondary benefits. Embedding analysis showed that class clusters became separable after semantic conditioning: the silhouette score on Edge-IIoTset reversed from negative 0.361 to positive 0.183, and silhouette scores correlated strongly with accuracy across all 24 dataset-stage combinations. Eight of fifteen attack classes that the baseline model completely failed to detect, scoring zero on the F1 metric, became detectable after preprocessing, though seven classes with extremely few training examples, some with as few as two samples, remained undetected, indicating that data preparation alone cannot solve the problem of class imbalance. Training also converged faster on heterogeneous datasets, with speedups of 3.3 to 5.8 times attributed to better-conditioned inputs.

The statistical evidence is strongest where replication was performed. Multi-seed validation on UNSW-NB15 multiclass classification, using five random seeds, showed a mean improvement of 8.9 percentage points with non-overlapping 95 percent confidence intervals, statistical significance at p below 0.01, and a large effect size with Cohen’s d above 0.87. MAFT also reduced seed sensitivity sixfold on that task. An extended multi-seed evaluation across all six datasets was conducted in binary classification mode, where seed sensitivity fell by up to 12.5 times on Edge-IIoTset, but the authors caution that this extension should be interpreted as seed-stability evidence rather than direct multiclass replication, and that the headline multiclass gains on ToN-IoT and Edge-IIoTset remain single-seed results.

The gains were not universal, and the authors are careful about this. In a comparison against eight alternative preprocessing methods, adaptive transformations such as PowerTransformer and QuantileTransformer achieved higher accuracy on some datasets, outperforming MAFT by more than 16 percentage points on Edge-IIoTset, though at the cost of per-column exception handling, greater computational overhead, and reduced interpretability. MAFT remained competitive elsewhere, within 0.2 to 0.8 percentage points of the best method on ToN-IoT, CICIDS2017, and UNSW-NB15, while producing valid outputs without exception handling and with preprocessing overhead under one second. Cross-model experiments showed that the benefit is specific to gradient-based architectures: the MLP and FT-Transformer improved on heterogeneous datasets, while Random Forest and XGBoost, which are scale-invariant by design, showed no benefit.

A further ablation clarified where the taxonomy itself matters. In binary classification, simply applying robust scaling uniformly to all features achieved results nearly identical to the full semantic grouping, indicating that outlier-resistant scaling is the dominant driver in that regime and that the taxonomy’s value there lies in interpretability and auditability rather than additional accuracy. The authors argue the six-group taxonomy becomes more relevant in multiclass settings with rare classes, where applying robust scaling to bounded categorical and protocol-specific features would be semantically inappropriate and could weaken minority-class signals.

The framework also depends on feature names following recognizable conventions; datasets with opaque column names rely more heavily on the statistical fallback, which may produce less meaningful groupings. All datasets in the study were generated in testbed environments, so production network traffic may exhibit distributional differences not captured by these benchmarks. The authors frame the observed relationship between heterogeneity and benefit as an empirically supported heuristic rather than a proven predictor, noting that validation on additional independent datasets would be needed before treating it as a rule.

The central lesson the authors draw is that data preparation should be treated as a measurable design choice in intrusion detection rather than a fixed implementation detail. Because MAFT operates entirely at the preprocessing level and leaves the downstream model unchanged, it can be adopted without retraining existing architectures or modifying deployed systems, an important property for regulated environments where model changes require recertification. Its deterministic, auditable mapping from feature type to scaling method makes it suitable for settings where pipeline decisions must be documented and justified, while the boundary conditions the study identifies allow practitioners to assess likely applicability from a feature-profile audit before committing to full model training.

Subject of Research: Modality aware feature taxonomy for robust AI driven intrusion detection across heterogeneous cyber physical data

Article Title: Modality aware feature taxonomy for robust AI driven intrusion detection across heterogeneous cyber physical data

Article References: Yadav, S. K., Verma, V. K., & Bhardwaj, V. (2026). Modality aware feature taxonomy for robust AI driven intrusion detection across heterogeneous cyber physical data. Discover Artificial Intelligence, 6(1), Article 1632. https://doi.org/10.1007/s44163-026-02053-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02053-x

Keywords: Modality, aware, feature, taxonomy, robust, driven, intrusion, detection, across, heterogeneous, cyber, physical