Hisham Mohammad

Hisham Mohammad

Auto-generated profile, based on publicly available information

Hisham Mohammad


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

15 August 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
AMAwwab Mohammad
Awwab Mohammad
HMHisham Mohammad
Hisham Mohammad
Federated learning is an attractive recipe for training network intrusion detectors across organisational boundaries, but most existing federated IDS proposals stop short of three things reviewers reasonably ask for: a working defence against client poisoning, a head-to-head comparison against modern federated baselines, and statistically significant accuracy gains. We address all three. We introduce FedXAI-IDS+, a two-stage federated pipeline that warms up with differentially-private FedAvg using a round-aware adaptive top-k schedule (k anneals from 0.50 to 0.10 over fifteen rounds), then hands off to a server-side LightGBM refinement protected by a Krum-style filter on the per-client summaries. Across NSL-KDD, CIC-IDS-2017, and TON_IoT — averaged over ten random seeds in an IID regime, five seeds per non-IID Dirichlet level, and three seeds per poisoning level — FedXAI-IDS+ reaches binary F1 of 0.9946, 0.9975, and 0.9987 respectively, beating FedAvg, FedProx, SCAFFOLD, and a federated XGBoost ensemble at p<0.001 by Wilcoxon signed-rank, while statistically tying centralised XGBoost and LightGBM. When two of five clients send label-flipped training data, FedXAI-IDS+ keeps F1 at 0.991, whereas SCAFFOLD drops to 0.46 and FedAvg to 0.78. Ablation isolates the LightGBM refinement as the load-bearing component and shows DP noise, top-k compression, and the Byzantine filter are accuracy-neutral under clean conditions — defences that cost nothing. Code and run artefacts are released.

15 August 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
AMAwwab Mohammad
Awwab Mohammad
HMHisham Mohammad
Hisham Mohammad
Security teams are increasingly unwilling to act on alerts from classifiers whose reasoning they cannot inspect, turning explainability into a hard requirement for network intrusion detection systems (NIDS). This paper proposes a lightweight feature-tokenized Transformer for multi-class intrusion detection, benchmarked on NSL-KDD (125,973 training and 22,544 test records across five categories, whose test partition contains 17 attack subtypes never seen during training) against Random Forest, an MLP, and an RBF-kernel SVM. It reaches 77.28% accuracy, weighted F1 of 0.723, and MCC of 0.670, edging out the other deep models while Random Forest keeps the lead on macro-precision (0.835) and macro AUC (0.957). SHAP gradient attributions for the Transformer are cross-checked against Random Forest's impurity-based importances, agreeing on 6 of the top 10 discriminative features but diverging on the rest — evidence feature-importance rankings are model-dependent rather than fixed. All models, including the proposed Transformer, initially failed to detect the minority R2L and U2R attack categories (recall 0.8% and 0%), reproducing a well-documented NSL-KDD limitation. Rather than stopping at diagnosis, we test two mitigations: inverse-frequency class-weighted loss and SMOTE-based minority oversampling. Classweighting lifts R2L and U2R recall to 29.7% and 40.3% at a cost to accuracy, while SMOTE lifts them to 25.5% and 28.4% while also improving accuracy (78.3%) and MCC (0.680) beyond baseline. These results show attention-based tabular Transformers are a viable, interpretable, improvable alternative to classical ensemble detectors, where explainability analysis can directly motivate a low-cost fix for a benchmark's best-known failure mode rather than merely describing it.

6 October 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
AMAwwab Mohammad
Awwab Mohammad
HMHisham Mohammad
Hisham Mohammad
Machine-learning intrusion detection for the Internet of Medical Things (IoMT) is evaluated almost entirely by point accuracy on a single benchmark, and published figures cluster above 97%. We show that part of that band is an artifact of how the benchmarks were labelled. In WUSTL-EHMS-2020 the source-MAC field records which machine the authors designated as the attacker; retaining it lifts a tuned gradient-boosted tree from 94.3% accuracy to 100.0% and its F1 from 0.800 to 1.000. We generalize this into a mechanism: an identifier leaks when the labelling procedure was defined in terms of it, not because of what kind of column it is. We release a detector that tests for this from the distributed data alone in seconds, and audit seven public IoT and IoMT benchmarks. Five ship endpoint addresses; two leak through them, and both come from one laboratory's capture pipeline. Three do not, at label purities of 0.836, 0.393 and 0.042, which is what makes the finding more than a tautology. The second leaking benchmark publishes a removal list naming six columns, of which our detector recovers three unaided; its IoMT counterpart publishes no such warning, and a screen of eight papers using it finds seven reporting at least 97% accuracy and none naming the columns they excluded. We then ask what the corrected task permits, building CMC-Net, which combines cross-modal attention, a consistency objective and an evidential head, as the instrument. Under repeated 5x5-fold cross-validation with corrected significance testing, it does not beat tuned XGBoost.