Adaptive AI framework keeps intrusion detection accurate as data drifts

A machine-learning framework that keeps its accuracy when the data underneath it changes over time has been reported in Frontiers in Artificial Intelligence. The system, called TEAL-XAI, was tested on three widely used network intrusion detection benchmarks and, according to its authors, held up far better than four standard baselines when a simulated attack on the data distribution was introduced mid-stream. The work addresses a well-known weakness of static classifiers: models trained on the assumption that data are independent and identically distributed slowly lose their footing as real-world traffic, fraud patterns, or health signals evolve, a phenomenon known as concept drift.

The researchers, from Vellore Institute of Technology in Tamil Nadu, India, built their framework around three components that operate together. The first is an exponential temporal decay weighting scheme, in which older training samples receive progressively less influence on the model. The weight of each observation is computed as an exponential function of its age, controlled by a decay parameter lambda. A systematic grid search over ten values of this parameter identified 0.05 as the optimum, giving the choice an empirical basis rather than leaving it arbitrary. At this setting and with the roughly 395,000-record KDD Cup 1999 training set, the framework retains an effective sample size of about 87,430 observations, concentrating influence on roughly the most recent 22 percent of the data while preserving informational contributions from the rest.

The second component is adaptive sliding-window retraining. The temporally ordered training set is divided into overlapping windows, each covering 40 percent of the data and advancing in steps of 10 percent. Within each window, the model is trained on a 90 percent sub-split and validated on the remaining 10 percent, and the decision threshold that separates normal from attack traffic is recalibrated by sweeping 81 candidate values to maximize the F1-score on that window’s validation data. The best-performing configuration is retained, and a final model is retrained on the complete training set with full temporal decay weights. This per-window threshold calibration matters because the class-conditional probability distributions shift across temporal epochs, so a single global threshold cannot stay well calibrated as the data evolve.

The third component is what the authors describe as a closed-loop explainability module. The framework uses SHAP, a method grounded in cooperative game theory that assigns each feature a theoretically consistent attribution for every prediction. Most applications of SHAP are retrospective: a model is trained, and the explanations annotate its output afterward. In TEAL-XAI, the global feature importances computed as mean absolute Shapley values are fed back into the training loop itself. After each retraining cycle, the top 15 features ranked by SHAP importance replace the original 20 features selected by mutual information, and the reduced set enters the XGBoost base learner with a column sampling rate of 0.85. The authors are careful to note that merely reordering columns would not change how XGBoost builds its trees, since the algorithm’s internal column-block structure is order-independent; the influence operates through feature selection, which guarantees that highly ranked features are included in the next adaptive cycle.

On the KDD Cup 1999 benchmark, the framework reached 99.73 percent accuracy, 99.71 percent F1-score, and an area under the receiver operating characteristic curve of 0.9998, the best scores among the five models tested on all five classification metrics. The baselines were Random Forest, XGBoost, LightGBM, and a four-layer deep neural network with 256, 128, 64, and 32 neurons. The framework outperformed the neural network by 0.35 percentage points on F1-score, which the authors interpret as evidence that explicit temporal adaptation contributes more than added network depth when data distributions are non-stationary. The same ranking order held on the NSL-KDD benchmark, where the framework achieved 99.41 percent F1-score and an AUC of 0.9996.

The most striking result concerns robustness to sudden concept drift. The evaluation simulated drift across ten sequential windows, with an abrupt change injected at Window 5 by stochastically flipping labels using Gaussian noise. Between the first and tenth windows, the framework’s accuracy fell by only 0.61 percentage points, from 99.73 to 99.12 percent. Random Forest lost 6.04 points, the deep neural network 6.41, XGBoost 5.50, and LightGBM 4.79. In relative terms, that makes the framework 9.9 times more resilient than Random Forest and 10.5 times more resilient than the neural network under this specific simulation. The authors translate the gap into operational terms: at Window 10, Random Forest’s 93.18 percent accuracy corresponds to a misclassification rate 7.8 times higher than the framework’s 0.88 percent, a difference they argue carries severe consequences for undetected intrusions.

The results on the third benchmark, UNSW-NB15, are more mixed, and the authors report this openly. This dataset, generated at the Australian Centre for Cyber Security with the IXIA PerfectStorm tool, reflects contemporary network protocols and post-2010 attack vectors across nine attack categories, making it a check against the legacy status of the two KDD-derived benchmarks. On its held-out test split, the framework attained 92.62 percent accuracy, 94.27 percent F1-score, and an AUC of 0.9843, tying the deep neural network for the highest F1-score among the five models and edging out XGBoost and LightGBM on that metric, though those two retained marginally higher AUC values. However, under five-fold stratified cross-validation the framework’s mean F1-score of 92.10 percent was lower than all four baselines, a difference confirmed as significant by the paired t-test and McNemar’s test. Under simulated sudden and gradual drift, all five models degraded by a comparable 50 to 53 percentage points, and the pronounced resilience advantage seen on the older benchmarks did not appear. The authors attribute this to the substantially larger training partition diluting the effective size of the recency-weighted window relative to the smaller datasets on which the decay parameter was originally validated, and they flag re-tuning for larger-scale benchmarks as an open direction rather than overstating generalization.

A six-configuration incremental ablation study on KDD Cup 1999 attributes the gains to each component in turn. Temporal weighting alone added 0.067 percentage points of F1-score over the baseline XGBoost configuration, adaptive retraining added a further 0.072, the closed-loop explainability module 0.042, temporal feature adaptation 0.044, and the final full configuration 0.039, for a cumulative improvement of 0.260 percentage points. The authors acknowledge that this absolute gain is modest given the near-ceiling performance of the baseline, but report that it remains statistically significant. The explainability contribution of 0.042 points is their evidence that the SHAP feedback loop is functionally beneficial rather than decorative.

One of the more novel analyses tracks how SHAP feature importances evolve across temporal windows. In the four windows before drift onset, the top three features by mean absolute SHAP value were src_bytes, count, and serror_rate. After the drift was injected, the importance of serror_rate rose by 47 percent in relative terms, reflecting the Gaussian perturbation’s amplification of error-rate statistics, while the importance of src_bytes declined by 23 percent. Critically, these importance shifts became detectable at Window 3, two windows before accuracy degradation became apparent at Window 5. The authors propose this as a proactive early-warning signal: the explanation layer can begin adapting feature priorities before measurable performance loss occurs, and it identifies which specific features are shifting, not merely that something has changed. They position this as complementary to recent label-free drift detectors that operate on classifier confidence or deep-learning representations but do not pinpoint the driving features.

Statistical validation used three independent tests across all four baseline comparisons: a paired t-test, a Wilcoxon signed-rank test, and McNemar’s test with Yates continuity correction, all at a significance level of 0.05, with p-values below 0.005 for the t-test and McNemar’s test on the primary benchmark. The authors also report limitations candidly. All three evaluation datasets come from simulated or curated intrusion detection environments, so cross-domain validation in fraud analytics or health monitoring remains future work. The drift scenarios were synthetically injected rather than observed in real streams. The deep learning baseline was a multilayer perceptron, leaving recurrent and attention-based architectures untested. Training time, at 38.52 seconds on the primary dataset, exceeded every baseline because of the multi-window retraining overhead. The framework’s base learner is XGBoost, which the authors argue suits tabular behavioral data under drift because it handles temporal dynamics through instance weighting rather than sequential state propagation. All experiments used a fixed random seed and standard library versions, and the three datasets are publicly available, supporting reproducibility of the reported results.

Subject of Research: An adaptive, explainable machine-learning framework for maintaining classification accuracy under concept drift in network intrusion detection.

Article Title: A temporal-evolving adaptive learning framework with integrated explainable AI dynamic behavioral pattern recognition

Article References: Hemachandran, K., & Vijay Anand, R. (2026). A temporal-evolving adaptive learning framework with integrated explainable AI dynamic behavioral pattern recognition. Frontiers in Artificial Intelligence, 9, 1929687. https://doi.org/10.3389/frai.2026.1929687

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1929687

Keywords: concept drift, adaptive learning, explainable AI, SHAP, network intrusion detection, XGBoost, temporal weighting, machine learning, TEAL-XAI, UNSW-NB15, KDD Cup 1999, drift detection