Contemporary web defense platforms often fail against previously unseen threat morphologies that bypass signature matching and standard machine learning models.
The framework, designated MutaShield-Net, combines software mutation testing principles with adversarial evolutionary learning to continuously stress-test detection boundaries.
- MutaShield-Net reached a 99.47% detection accuracy, a 98.92% F1-score, and a 99.21% AUC-ROC during experiments on the CIC-IDS2017 and CSE-CIC-IDS2018 benchmark datasets.
- The architecture reduced the false positive rate by 40.4% compared to the strongest single baseline, FL-WGAN-IDS, moving down from 0.52% to 0.31%.
- System design integrates three components: the Spectral Mutation Operator Engine, the Cross-Domain Attention-Gated Recurrent Transformer, and the Adversarial Mutation Fitness Evolutionary Loop.
The system operates through three interdependent components designed to generate traffic mutations and test detection classifiers simultaneously. The Spectral Mutation Operator Engine produces semantically valid traffic mutations across seven taxonomic attack families.
These generated mutations feed into the Cross-Domain Attention-Gated Recurrent Transformer. This detection backbone fuses temporal recurrence with multi-head cross-attention over raw packet features and application-layer semantic embeddings.
Mutant generators and detection classifiers engage in a min-max optimization game governed by a novel mutation survival score metric. This co-evolutionary loop, known as the Adversarial Mutation Fitness Evolutionary Loop, proactively stress-tests system boundaries before deployment against live traffic.
Existing approaches lack this capacity to evolve defensive models against novel attack variants. Extensive experiments demonstrated that the combined framework surpasses traditional machine learning configurations that rely solely on historical signature databases.
| Performance Metric | MutaShield-Net | FL-WGAN-IDS Baseline |
|---|---|---|
| Detection Accuracy | 99.47% | Not Reported |
| F1-Score | 98.92% | Not Reported |
| AUC-ROC | 99.21% | Not Reported |
| False Positive Rate | 0.31% | 0.52% |