CEDF-CS: Class-Balanced Prototype Condensationfor Resource-Efficient and Leak-Free Intrusion Detection in Industrial IoT and Enterprise Networks

Technical Summary: CEDF-CS: Class-Balanced Prototype Condensation for Resource-Efficient and Leak-Free Intrusion Detection

Problem Statement
Modern intrusion detection systems (IDS) face two critical challenges when deployed on resource-constrained Industrial IoT (IIoT) and edge devices. First, standard benchmarks like CICIDS2017 and WUSTL-IIoT-2021 contain millions of network flows with extreme class imbalance, where "Benign" traffic dominates and rare attack families are underrepresented, biasing classifiers toward majority classes. Second, existing dataset condensation methods—specifically recent "hierarchical fusion" approaches that use simple within-class record averaging—suffer from severe evaluation leakage. When averaging is performed on the entire dataset before splitting into train/test sets, the resulting synthetic training records are arithmetic combinations of records that also appear in the test set. This violates the independence assumption, artificially inflating performance metrics and failing to generalize to unseen data.

Methodology
The authors propose CEDF-CS, a framework that redesigns dataset condensation to be both resource-efficient and strictly leak-free. The methodology consists of three core components:

1. Leak-Free Evaluation Protocol: The study enforces a rigorous protocol where the dataset is split into stratified train/test partitions before any condensation occurs. All preprocessing (scaling, feature scoring) and condensation are applied exclusively to the training partition. The model is then evaluated on held-out, original (non-condensed) test records. This eliminates the train/test overlap inherent in previous "hierarchical fusion" studies.
2. Class-Balanced Prototype Condensation: Instead of naive averaging, the framework generates synthetic prototypes using k-means clustering. Crucially, it introduces a class-balanced budget allocation. Rather than allocating the compression budget proportional to the original class sizes (which would further marginalize rare attacks), the budget is redistributed equally across all classes. This ensures that rare attack families receive a proportional number of synthetic prototypes, preserving their representational capacity within the compressed set.
3. CPU-Only Classifier Suite: To ensure deployability on edge hardware, the framework utilizes three CPU-native classifiers: a shallow Multilayer Perceptron (MLP), an RBF-Support Vector Machine (SVM), and a Histogram Gradient Boosting (HGB) classifier. No GPU acceleration is required.

The framework evaluates seven variants, including the original "avg-fusion" baseline (which collapses under the leak-free protocol), class centroids, stratified random subsampling, k-means prototypes (unbalanced), and the proposed class-balanced k-means prototypes (Variant F). A per-class Gaussian Mixture Model (GMM) synthesis variant is also tested to preserve within-class covariance.

Key Results
The study was conducted over five random seeds on two benchmarks: WUSTL-IIoT-2021 (binary, 1.19M flows) and CICIDS2017 (8-class, 2.83M flows).

• Failure of Previous Methods: Under the leak-free protocol, the original "avg-fusion" method (Variant B) suffered catastrophic performance degradation. On WUSTL-IIoT, the F1 score for SVM dropped from ~0.99 (in leaky protocols) to 0.17. On CICIDS2017, macro-F1 for the neural network dropped from 0.67 to 0.31. This confirms that previous high-performance claims were artifacts of evaluation leakage.
• Success of CEDF-CS (Variant F): The class-balanced prototype scheme achieved performance indistinguishable from full-data training while using only 1/32 of the training records (effective compression up to 512× on WUSTL-IIoT due to binary class balancing).
  • WUSTL-IIoT: Variant F achieved an F1 of 0.996 and Balanced Accuracy of 0.9996 with HGB, matching the full-data baseline.
  • CICIDS2017: Variant F significantly improved Balanced Accuracy (a critical metric for rare class detection) over full-data training: 0.838 (NN) and 0.843 (HGB) compared to 0.659 for full-data.
• Metric Trade-offs: While Variant F maximizes Balanced Accuracy and per-class recall for rare attacks, the unbalanced k-means variant (Variant E) achieved the highest Macro-F1 (0.699 vs. 0.671 for full-data NN). This highlights a trade-off: Variant F sacrifices some precision on the dominant "Benign" class to ensure rare attacks are detected, whereas Variant E maintains a better balance of precision and recall across all classes.
• Mechanism of Failure: The paper provides an "association-amplification" analysis showing that within-class averaging artificially inflates feature-label associations in the training distribution (due to standard error shrinkage), leading to overfitting that does not transfer to held-out data.

Significance and Claims
The paper claims that dataset condensation can be successfully adapted for intrusion detection if two conditions are met: the elimination of evaluation leakage and the explicit handling of class imbalance through budget reallocation.

• Operational Relevance: CEDF-CS demonstrates that high-performance IDS can be trained on edge devices using only CPU resources, reducing training data requirements by orders of magnitude without sacrificing detection capability on rare attacks.
• Reproducibility and Rigor: The study serves as a corrective to the literature, demonstrating that previously reported "spectacular" gains from simple linear condensation were largely due to protocol flaws. It establishes a new standard for evaluating condensation methods on imbalanced, large-scale tabular data.
• Generalizability: The authors posit that the mechanism of preserving multi-modal class structure via k-means prototypes and redistributing compression budgets is a generic solution for long-tailed tabular classification, applicable beyond IDS to domains like fraud detection and medical logs, though this remains a hypothesis.

The paper concludes that class-balanced prototype condensation is a viable, deployment-ready strategy for resource-efficient intrusion detection, offering a practical alternative to deep learning models that require GPUs and massive datasets.