PRAISE-IoMT: A Privacy-Aware, Adversarially Robust, and Explainable CNN-SNN Framework for Cross-Domain IoT-to-IoMT Intrusion Detection

# PRAISE-IoMT Reproducibility Package This package is the canonical computational companion for the PRAISE-IoMT manuscript. It centralizes dataset provenance, the executable 18-feature representation, group-aware partitions, strict source-only transfer views, model configurations, checkpoints/model artifacts, optimizer settings, adversarial constraints, calibration, explainability, resource profiling, metric generation, and manuscript crosswalks. The package implements three transfer-evaluation views in both directions: 1. **Natural unique view** — all unique target observations under the target domain's original class distribution.2. **Balanced unique view** — equal class support using unique target observations without replacement.3. **Resampled balanced sensitivity view** — the 6,000/6,000 construction retained only as a sensitivity analysis and never interpreted as 12,000 independent raw observations. The `CENTRAL_REGISTRY.xlsx` workbook is the sole status register. Manuscript-facing text and package documentation remain free of workflow-status language. ## Reproducibility levels **Exact evaluation reproduction:** dataset + manifest + preprocessing artifact + checkpoint/model artifact + evaluation script -> predictions -> metrics -> manuscript table values. **Training reproduction:** dataset + group split + feature map + configuration + seed + training/search code -> checkpoint/model artifact -> evaluation outputs. ## Main entry points - `scripts/run_pipeline.py` — orchestrates reproducibility stages.- `transfer/build_transfer_views.py` — creates natural, balanced-unique, and resampled-balanced target manifests.- `models/cnn_snn_model.py` — PRAISE-IoMT CNN-LIF-SNN architecture.- `models/train_source.py` — source-domain training and adversarial fine-tuning.- `adversarial/attacks.py` — constrained FGSM/PGD.- `adversarial/pgd_convergence_check.py` — PGD-10/5 versus PGD-50/10 evaluation.- `metrics/generate_metrics.py` — predictive, transfer, and adversarial metrics.- `scripts/reproduce_from_checkpoint.py` — regenerates predictions and metrics from a saved checkpoint.- `optimization/search_driver.py` — fixed-budget two-stage BWO-BKA search and Pareto archive management. ## Artifact identity Every model artifact, manifest, configuration, and result file has a stable identifier. The final release manifest records SHA-256 checksums so that manuscript values can be traced to the exact computational artifacts used to generate them. ## Canonical data and feature layer The common network representation is `COMMON18-ARGUS-V1`. WUSTL-EHMS-2020 contributes the corresponding native ARGUS flow fields directly from the official CSV; mean packet size is calculated as total bytes divided by total packets. MedBIoT uses the official fine-grained raw PCAP captures and the matching ARGUS/ra fields, so source and target are aligned by flow semantics and units rather than column-name similarity. WUSTL source-domain grouping uses 40 non-overlapping ordered acquisition blocks over the stable `Packet_num` sequence. The row-level and group-level manifests preserve raw-row identity, class composition, and partition membership. MedBIoT grouping uses resolved device identity as the primary key with capture, campaign-stage, and file-hash fallbacks. The package retains the official raw WUSTL CSV, a processed WUSTL common-18 dataset, native-schema documentation, raw-row SHA-256 identifiers, MedBIoT official file inventory, and exact partition/transfer manifests. ## Model, training, optimization, and checkpoint layer The canonical model implementation is `PRAISE-CNN-SNN-V1`, with 829,057 trainable parameters verified by runtime audit. The source-training protocol stores both pre-adversarial and final checkpoints for each source domain and seed. Every checkpoint carries configuration, feature-schema, partition-manifest, threshold, calibration, parameter-count, framework-version, and package-integrity metadata. The optimizer is an explicitly executable two-stage Pareto search: BWO operators are used for generations 1–20, followed by a Pareto-archive adaptation of the published BKA attack and migration equations for generations 21–40. The fixed budget is 30 candidates over 40 generations, or 1,200 candidate evaluations per seed. Search objectives are FNR, FPR, normalized inference-work proxy, and normalized parameter count. Measured Jetson energy remains a post-selection measurement. The executable baseline set contains CNN-only, SNN-only, XGBoost, Random Forest, and SVM. Logistic regression is not part of the baseline artifact set. ## Evaluation and metric-generation layer `PRAISE-EVAL-V1` standardizes within-domain, strict-transfer, adversarial, calibration, explainability, and resource evaluation. Strict transfer is evaluated in both directions under the natural-unique primary view, balanced-unique secondary view, and resampled-balanced sensitivity view. CNN-only and XGBoost provide representative source-only transfer comparisons under the same preprocessing and threshold rules. Stochastic adversarial evaluation uses expectation-over-transformation gradients across Bernoulli spike realizations. The principal PGD configuration uses 10 iterations and 5 restarts; the convergence check uses 50 iterations and 10 restarts at epsilon 0.05. ASR is computed only over attack observations that were correctly detected before perturbation. Calibration uses scalar-logit binary temperature scaling and reports Brier score and 15-bin ECE on within-domain held-out data. Resource profiling includes median, P95, P99, run variability, idle-subtracted energy, peak memory, and layer-wise spike activity for classifier inference on the evaluated Jetson configuration. ## Numerical control and manuscript synchronization layer The numerical layer contains seed-level within-domain and transfer tables, count-consistent transfer confusion matrices, natural/unique and balanced-unique transfer summaries, the resampled sensitivity results, stronger-PGD convergence values, calibration values, extended resource metrics, configuration-comparison values, and the manuscript numeric crosswalk. A numerical integrity audit checks every transfer confusion matrix against the reported accuracy, macro-F1, attack recall, and FPR; verifies recall/FNR complementarity; and checks latency ordering. Published WUSTL-EHMS-2020 and MedBIoT studies are included only as a plausibility envelope because their tasks, splits, and preprocessing are not directly comparable. ## Manuscript and reviewer-response synchronization The release includes the minimally revised manuscript and the point-by-point reviewer response under `release/manuscript/`. Manuscript structure is retained except where a reviewer-requested scientific or reproducibility correction requires a local replacement. The manuscript uses red text for revisions and yellow highlighting for registry-controlled numerical values. Cross-domain transfer, adversarial denominators, feature alignment, optimizer specification, calibration, resource profiling, and claim boundaries are synchronized with the package artifacts and numeric crosswalk.