IntrusionX: Hybrid Deep Learning IDS

• The paper presents IntrusionX, a hybrid deep learning framework that combines Conv1D layers for spatial feature extraction with LSTM for temporal modeling in IDS.
• It incorporates an imbalance-aware training pipeline with stratified splitting and dynamic class weighting to significantly improve minority class detection, particularly for U2R and R2L attacks.
• The Squirrel Search Algorithm is used to optimize key hyperparameters, ensuring a reproducible and high-performance model on benchmarks like NSL-KDD.

IntrusionX is a hybrid deep learning framework for network intrusion detection that integrates Convolutional Neural Networks (CNNs) for local feature extraction and Long Short-Term Memory (LSTM) networks for temporal modeling, with the architecture further optimized using the Squirrel Search Algorithm (SSA) for hyperparameter tuning (Farabi et al., 1 Oct 2025). It is presented against a problem setting defined by evolving cyberattacks, high-dimensional traffic data, and severe class imbalance in benchmark datasets such as NSL-KDD. Its reported novelty is a reproducible, imbalance-aware design with metaheuristic optimization, and its evaluation emphasizes not only global accuracy but also rare-class detection, especially the minority classes U2R and R2L (Farabi et al., 1 Oct 2025).

1. Problem setting and design objective

IntrusionX is designed for network intrusion detection in conditions where benchmark traffic data are both high-dimensional and severely imbalanced. The framework is explicitly motivated by the observation that Intrusion Detection Systems face persistent challenges due to evolving cyberattacks, high-dimensional traffic data, and severe class imbalance in benchmark datasets such as NSL-KDD (Farabi et al., 1 Oct 2025). In this formulation, the central modeling problem is not only to distinguish normal from malicious traffic, but also to preserve detection quality for rare attack classes that are often underrepresented during training.

The paper situates rare-class detection as a first-order objective rather than a secondary evaluation criterion. In the five-class setting—Normal, DoS, Probe, R2L, and U2R—the emphasis on minority class recall reflects the operational importance of minimizing false negatives for rare but critical attacks. This framing aligns with broader IDS research showing that many traditional and prior deep learning IDSs perform well on majority classes but degrade sharply on minority or unstable attack behaviors (Farabi et al., 1 Oct 2025). Related anomaly-detection studies likewise report that attacks with unstable, distributed, or non-repeatable behavior as Fuzzing, Worms and Botnets are more difficult to detect, even when strong unsupervised baselines such as Isolation Forests, One-Class Support Vector Machines, and Self-Organizing Maps are used (Zoppi et al., 2020).

2. Hybrid Conv-LSTM architecture

The core architectural decision in IntrusionX is the coupling of Conv1D layers with LSTM layers. The Conv1D layers extract local spatial correlations in input features, including relationships among NSL-KDD’s 41 features, helping reduce redundancy and focus on informative local patterns in network traffic. The LSTM layers then model temporal dependencies and sequential traffic characteristics, learning how events unfold over time. Dense layer(s) map the resulting representations to output classes for binary or 5-class classification (Farabi et al., 1 Oct 2025).

The integration is strictly sequential: the Conv1D layers feed into the LSTM layer(s), so that refined spatial features are subsequently interpreted for temporal relationships. In the paper’s summary notation, the model layers are described as:

$\omega_{c(i)}$0

This arrangement distinguishes IntrusionX from CNN-only and LSTM-only formulations discussed in prior IDS literature. The paper states that traditional and even prior deep learning IDSs, including CNN or LSTM-only approaches, typically excel on majority classes but perform very poorly on U2R and R2L, sometimes near 0% recall (Farabi et al., 1 Oct 2025). A plausible implication is that the hybridization is intended to capture both local feature structure and sequential dependence in a way that is specifically beneficial for minority-class separability.

The architecture remains a black-box deep model in the sense described by explainable IDS surveys, where deep neural networks, random forests, and SVMs are treated as powerful but opaque predictors that require post-hoc explanations if interpretability is required (Neupane et al., 2022). IntrusionX, as summarized, does not include an explainer module.

3. Preprocessing, stratified splitting, and imbalance mitigation

IntrusionX incorporates a preprocessing and training pipeline that is explicitly imbalance-aware. The pipeline stages are: data preprocessing, stratified splitting, dynamic class weighting, Conv1D + LSTM modeling, SSA-based hyperparameter optimization, and evaluation with emphasis on rare-class detection (Farabi et al., 1 Oct 2025).

The preprocessing stage includes three concrete operations. First, categorical features—protocol type, service, and flag—are one-hot encoded, transforming them into binary vectors. Second, continuous features are normalized using Min-Max scaling to ensure equitable input ranges. Third, data cleaning ensures the processed data is suitable for learning, free from anomalies or errors (Farabi et al., 1 Oct 2025). These steps define the input representation supplied to the neural architecture.

The train/validation/test partition is created using stratification, ensuring class distributions are preserved across all subsets. In imbalanced intrusion datasets, this matters because rare classes might otherwise be missing from validation or test splits. The paper presents this as a leak-free splitting strategy and includes it among the framework’s novelty claims (Farabi et al., 1 Oct 2025).

Dynamic class weighting is incorporated directly into the categorical cross-entropy loss function:

$$\mathcal{L}_{weighted} = -\sum_{i=1}^N \omega_{c(i)} y_i \log(\hat{y}_i)$$

where $\omega_{c(i)}$ is the weight for the true class of instance $i$, $y_i$ is the true label, and $\hat{y}_i$ is the predicted probability. The summary further notes that $\omega_{c(i)} = \frac{N}{N_{c(i)}}$ is typically used, with $N$ total samples and $N_{c(i)}$ samples of class $c(i)$ (Farabi et al., 1 Oct 2025).

The operational consequence is explicit: misclassification of rare classes such as U2R and R2L is penalized more heavily, increasing their recall and shifting the learning focus from only maximizing overall accuracy to also safeguarding against overlooking infrequent attacks. This is a materially different strategy from treating class imbalance solely through data sampling; the paper describes the framework as an explicit, reproducible, and thorough imbalance-management pipeline, not just data sampling but weighting and leak-free splitting (Farabi et al., 1 Oct 2025).

4. Squirrel Search Algorithm and hyperparameter optimization

IntrusionX uses the Squirrel Search Algorithm as a metaheuristic hyperparameter optimization mechanism. SSA is described as being inspired by the dynamic foraging behavior of squirrels in nature and is used to optimize critical model hyperparameters, including the number of convolutional filters, the number of LSTM units, and the learning rate, among others (Farabi et al., 1 Oct 2025).

The reported role of SSA is threefold. It is said to ensure more effective exploration of the hyperparameter space compared to manual or grid search; to balance accuracy, minority class recall, and computational efficiency by adapting to optimal parameter configurations depending on dataset characteristics; and to produce a model better tuned for both performance and resource requirements (Farabi et al., 1 Oct 2025). The workflow is summarized in three steps: each “squirrel” represents a set of model hyperparameters; fitness is determined by the model’s performance on validation accuracy or recall; and the population evolves by mimicking squirrel foraging behaviors, adaptively moving toward more optimal parameter sets.

The paper identifies this integration as a novelty claim: IntrusionX is described as the first to combine SSA-driven hyperparameter tuning with a Conv-LSTM network specifically for intrusion detection (Farabi et al., 1 Oct 2025). In encyclopedic terms, this places SSA at the level of model-selection policy rather than feature engineering or inference-time defense. By contrast, other IDS papers in the surrounding literature focus on different optimization targets: autonomic response selection via Theory of Expected Utility in distributed computing (Vieira et al., 2018), or SHAP-based adversarial detection and zero-touch mitigation in O-RAN environments (Paltun et al., 7 Mar 2025). IntrusionX’s optimization target is the neural architecture and its training dynamics on NSL-KDD.

5. Experimental evaluation and reported performance

The experimental evaluation reported for IntrusionX is conducted on NSL-KDD for binary classification and 5-class classification (Farabi et al., 1 Oct 2025). The binary setting distinguishes Normal vs. Attack; the five-class setting distinguishes Normal, DoS, Probe, R2L, and U2R.

The key results are summarized below.

Task	Accuracy	Key class metrics
Binary	98%	Precision & Recall: >97%; ROC-AUC: 0.9986
Five-Class	87%	Weighted F1 Score: 0.90; R2L Recall: 93%; U2R Recall: 71%

For binary classification, the framework reports 98% accuracy, Precision & Recall above 97%, and ROC-AUC of 0.9986 (Farabi et al., 1 Oct 2025). For five-class classification, it reports 87% accuracy and a Weighted F1 Score of 0.90, with Majority Class Recall above 90% for Normal, DoS, and Probe, and Minority Class Recall of 93% for R2L and 71% for U2R (Farabi et al., 1 Oct 2025).

The rare-class results are central to the framework’s stated contribution. The paper characterizes the 71% recall for U2R and 93% for R2L as sizable improvements over most prior work, where such classes are often under 50% recall (Farabi et al., 1 Oct 2025). It further states that hybrid models in the literature often lack either optimization or explicit imbalance handling and hence suffer on minority classes. The evaluation is therefore used not only to report aggregate performance but also to support the claim that the framework’s hybrid Conv-LSTM learning, automated SSA optimization, and rigorous imbalance mitigation act jointly.

A common misconception in reading IDS metrics is to treat binary accuracy as sufficient evidence of deployment readiness. IntrusionX’s own evaluation structure argues against that simplification: the paper foregrounds five-class performance and minority-class recall, indicating that aggregate binary discrimination can obscure failure modes on rare attack categories (Farabi et al., 1 Oct 2025).

6. Position within intrusion-detection research

IntrusionX belongs to a broader IDS landscape that includes signature-based response systems, anomaly detectors, explainable IDS, and privacy-preserving collaborative architectures. Its contribution is specific: a hybrid Conv-LSTM detector with SSA-driven hyperparameter tuning and explicit imbalance handling for NSL-KDD (Farabi et al., 1 Oct 2025).

This positioning becomes clearer when compared with adjacent research directions. In distributed computing, the Autonomic Intrusion Response System (SARI) uses autonomic computing via the Monitor-Analyse-Plan-Execute-Knowledge model, MapReduce for large-scale processing of log and network data, and Theory of Expected Utility for response selection and self-healing (Vieira et al., 2018). IntrusionX does not implement MAPE-K, MapReduce-based detection, or utility-based response planning; its scope is network intrusion detection through deep representation learning.

In explainable and adversarially robust IDS research, SHAP has been used as a defense mechanism rather than only an interpretation layer. A robust IDS for the O-RAN RRC layer integrates an auto-encoder neural network with SHAP-based outlier detection on explanation distributions and supports zero-touch mitigation, reporting SHAP accuracy of 0.93, precision of 0.76, recall of 1.00, and F1-score of 0.87 under an AE-BIM scenario (Paltun et al., 7 Mar 2025). More generally, the survey literature on X-IDS identifies a trade-off between high predictive power and explanation quality, and proposes a generic architecture with pre-modeling, modeling, and post-modeling phases plus human-in-the-loop feedback (Neupane et al., 2022). IntrusionX, as summarized, is not an X-IDS and does not provide SHAP, LIME, LRP, or counterfactual modules.

The framework also sits alongside complementary operational concerns that it does not directly solve. Cloud monitoring research emphasizes the credibility of log data through MAC chains and reconstructability through Shamir’s $(k,n)$ threshold secret sharing (Weir et al., 2024). Privacy-preserving collaborative IDS architectures emphasize PIR for threat-intelligence lookup and SMC for cross-organization telemetry analytics (Dara et al., 2016). A plausible implication is that an operational deployment around IntrusionX could combine its detector with these monitoring, integrity, or collaboration mechanisms without altering its core Conv-LSTM + SSA design.

7. Significance, limitations, and likely extensions

The significance claimed for IntrusionX lies in the combination of four elements: hybrid Conv-LSTM learning, automated SSA optimization, rigorous imbalance mitigation, and public, reproducible implementation (Farabi et al., 1 Oct 2025). The framework is explicitly described as reproducible and imbalance-aware, and its reported public code and implementation reinforce that orientation.

Its limitations are also implicit in the scope of the reported study. The evaluation described in the summary is on NSL-KDD, and the performance claims are tied to binary and five-class classification within that benchmark (Farabi et al., 1 Oct 2025). A plausible implication is that transfer to other environments—such as hybrid cloud-edge infrastructures, O-RAN RRC-layer monitoring, or collaborative cross-organization analytics—would require additional validation under those settings. This caution is consistent with the broader IDS literature, where dataset characteristics, feature discriminativeness, and attack heterogeneity materially affect performance (Zoppi et al., 2020).

A second likely extension concerns interpretability and response orchestration. The explainable IDS literature argues that deep models are often treated as black box models and do not provide a justification for their predictions, creating a barrier for analysts who must make decisions about detected threats (Neupane et al., 2022). Separate work on robust IDS shows that explanation distributions can be used not only for interpretability but also for adversarial detection and zero-touch mitigation (Paltun et al., 7 Mar 2025). This suggests a possible future direction in which IntrusionX’s Conv-LSTM detector is paired with an explainer module or mitigation layer, although such a component is not part of the framework as summarized in the cited paper.

Within the deep-learning IDS literature, IntrusionX is therefore most precisely understood as a benchmark-focused, imbalance-aware hybrid detector whose central claim is not merely high overall accuracy, but improved rare-class recall through coordinated architectural design, stratified evaluation, weighted loss, and SSA-based model selection (Farabi et al., 1 Oct 2025).