What question did this study set out to answer?

This research aims to enhance cybersecurity threat detection through a novel tri-component loss function in GANs for network traffic augmentation.

April 14, 2026Open Access

Enhanced cybersecurity threat detection using novel tri-metaheuristic loss functions in generative adversarial networks with adaptive attention preservation for network traffic augmentation

HKHeba M. KhalilBenha University AEAhmed ElrefaiyZagazig University MEMostafa ElbazKafrelsheikh University

Key Points

This research aims to enhance cybersecurity threat detection through a novel tri-component loss function in GANs for network traffic augmentation.
Developed a tri-component loss function framework for GANs
Integrated nine differentiable loss components including attention-based weighting and distribution alignment
Implemented energy-aware adaptive attention for resource allocation based on threat likelihood
Evaluated framework performance across seven cybersecurity datasets
Conducted ablation analysis to assess improvements from augmentation and loss combination
Achieved 98.73% accuracy and a 0.987 F1-score across datasets
Demonstrated a 40% reduction in training energy consumption
Found 49.4% improvement from class imbalance addressing and 50.6% from loss combination
Cross-dataset accuracy between 87.45% and 94.23% without retraining
Achieved 95.67% adversarial robustness under a perturbation budget of ε = 0.3

Abstract

This paper proposes a tri-component loss function framework integrated within Generative Adversarial Networks for network traffic augmentation in cybersecurity threat detection. The framework combines nine differentiable loss components: feature importance preservation via attention-based weighting, distribution alignment via Wasserstein distance, gradient regularization via gradient penalty, adversarial discrimination via hinge loss, embedding clustering via triplet constraints, curriculum scheduling via progressive difficulty adjustment, perturbation-aware training via projected gradient descent, multi-scale consistency via wavelet transform, and diversity promotion via cosine similarity regularization. We clarify that these components employ established techniques, with our contribution lying in their systematic integration and domain-specific adaptation rather than fundamentally new algorithms. Energy-aware adaptive attention dynamically allocates computational resources based on threat likelihood, reducing training energy consumption by 40% (76.8 kWh versus 128.4 kWh baseline). Experimental evaluation across seven cybersecurity datasets (NSL-KDD, UNSW-NB15, CIC-IDS2017, CIC-IDS2018, Bot-IoT, CICDDOS2019, CSE-CIC-IDS2018) yielded 98.73% accuracy and 0.987 F1-score. Ablation analysis revealed that 49.4% of improvement stems from addressing class imbalance through augmentation, while 50.6% derives from the proposed loss combination, with 2.0% additional synergistic benefit. Cross-dataset transfer achieved 87.45–94.23% accuracy without retraining. Adversarial robustness evaluation of 95.67% accuracy under perturbation budget ε = 0.3. Limitations include poor infiltration attack detection (16.44–28.13% recall) and ground truth verification covering only 1.8% of deployment samples. Statistical significance was confirmed with p-values below 0.0001 and Cohen’s d exceeding 3.4. The framework provides evidence that systematic integration of established techniques with domain-specific adaptation can yield measurable improvements in cybersecurity applications under the evaluated conditions. Generalization to broader deployment contexts warrants further investigation.

AI에게 질문

Bookmark

View Full Paper