Malware Detection Eval

malware-detection-eval

Routing-Aware Explanations for Mixture of Experts Graph Models in Malware Detection — Shokouhinejad et al. (2026) (arXiv:2602.19025, 2026)

What this evaluates

Binary classification of software binaries as benign or malicious based on their control flow graphs. It probes the model's ability to learn graph-structured representations and route them through specialized experts for accurate detection.

Datasets

• BODMAS — total 122; splits: train (-1), test (-1)
• DikeDataset — total 319; splits: train (-1), test (-1)
• PMML — total 390; splits: train (-1), test (-1)

Metrics

• Accuracy (primary) — range: [0, 1]
  • Proportion of correctly classified samples out of the total number of samples. Calculated as (TP + TN) / Total.
• F1 — range: [0, 1]
  • Harmonic mean of precision and recall. Calculated as 2 * (Precision * Recall) / (Precision + Recall). Reported separately for benign and malicious classes.

Input / output format

Input: Control flow graphs (CFGs) extracted from binaries using angr, with node features compressed to 64 dimensions via a symmetric autoencoder.

Output: Binary classification label (Benign or Malware) per graph, plus routing gate weights and expert selections for MoE variants.

Scoring recipe

def compute_metrics(y_true, y_pred):
    correct = sum(1 for t, p in zip(y_true, y_pred) if t == p)
    accuracy = correct / len(y_true)
    tp = sum(1 for t, p in zip(y_true, y_pred) if t == 1 and p == 1)
    fp = sum(1 for t, p in zip(y_true, y_pred) if t == 0 and p == 1)
    fn = sum(1 for t, p in zip(y_true, y_pred) if t == 1 and p == 0)
    precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
    recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
    f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
    return {'accuracy': accuracy, 'f1': f1}

Common pitfalls

• Datasets have highly imbalanced class distributions (only DikeDataset is benign, BODMAS/PMML are malware), requiring careful stratified splitting to preserve class proportions.
• Node features are heavily preprocessed (439D to 64D via autoencoder); evaluating raw features without this compression will yield different results.
• Explainability metrics (Fidelity+/Fidelity-) are evaluated across sparsity levels (5%-95%), not just a single threshold.

Evidence (verbatim from paper)

The dataset was stratified and split into 80% for training and 20% for testing, preserving equal class proportions for benign and malicious samples in both subsets. Table[2] reports classwise precision, recall, F1, and overall accuracy for all models.

Citation

@misc{shokouhinejad2026routing,
  title={Routing-Aware Explanations for Mixture of Experts Graph Models in Malware Detection},
  author={Shokouhinejad et al. (2026)},
  year={2026},
  note={arXiv:2602.19025}
}

• arXiv: 2602.19025