Computational Intelligence Искусственный интеллект

Federated Attention‐Augmented Hybrid Framework for Real‐Time Intrusion Detection in Cloud Environments: Integrating Transformer‐Enhanced Convolutional Networks With Privacy‐Preserving Stochastic Optimization

V. Ramu, S. P. Santhoshkumar, Ramu Vankudoth, Eugene Berna Ignasimuthu, N. Naveen, A. Athiraja
Wiley · 2026-10-01 · DOI 10.1111/coin.70298

Аннотация

ABSTRACT The widespread adoption of cloud services in government and commercial environments has enlarged the attack surface exposed to advanced persistent threats, zero‐day exploits, and polymorphic distributed denial‐of‐service attacks. This paper presents the Federated Attention‐Augmented Hybrid Intrusion Detection Framework (FAAHIDF), which couples four established components through one training and deployment protocol: a Multi‐Head Spatial Attention Convolutional Encoder (MHSACE), a Federated Differential‐Privacy Stochastic Gradient Descent Optimizer (FDP‐SGDO), a Cross‐Domain Normalization and Augmentation Engine (CDNAE), and a Dynamic Threshold Calibration Module (DTCM). The contribution is therefore the coordinated interaction among local–global representation learning, privacy‐bounded federated updating, target‐aware domain alignment, and drift‐responsive decision calibration, rather than novelty of each primitive in isolation. CSE‐CIC‐IDS2018 is used as the labeled source benchmark. For UNSW‐NB15, only an unlabelled target‐development partition is exposed to CDNAE, whereas the held‐out target evaluation subset is excluded from normalization fitting, augmentation, domain alignment, threshold selection, and hyperparameter tuning; this experiment is consequently reported as target‐aware unsupervised domain adaptation rather than a fully unseen external test. Under the fixed experimental configuration, FAAHIDF records 99.87% accuracy, a 0.02% false‐positive rate, 0.9998 AUC, and 0.9981 macro‐F1 on CSE‐CIC‐IDS2018, and 98.74% accuracy with 0.9841 macro‐F1 on the held‐out UNSW‐NB15 target subset. These values are treated as point estimates. The revised manuscript limits statistical and state‐of‐the‐art claims to the available evidence, clarifies the cumulative privacy configuration (epsilon = 2.50, delta = 1e‐5), and adds module‐wise computational, communication, leakage‐control, baseline‐scope, and reproducibility analyses.

Читать у издателя →