Year: 2025
Type: article
Abstract: The increasing sophistication of cyber threats demands more advanced detection systems capable of identifying novel attacks while maintaining efficiency and interpretability. This paper presents a hybrid Transformer-LSTM neural network for real-time cyber threat detection through security event profiling. Unlike existing approaches, our model combines the long-range dependency capture of Transformers with the temporal pattern recognition of LSTMs, enabling more accurate identification of complex attack sequences. Additionally, we integrate a self-supervised learning mechanism that allows the model to dynamically adapt to emerging threats without requiring retraining on fixed datasets. Evaluated on the CIC-IDS2017 dataset, our approach achieves 98.7% precision, 97.5% recall, and a 98.1% F1-score, outperforming state-of-the-art methods such as DeepLog, LSTM, and GNN-based detectors. The model also demonstrates computational efficiency, with a 30% reduction in training time compared to existing architectures. Beyond performance improvements, our framework incorporates attention-based explainability, providing security analysts with interpretable insights into detection decisions. These advancements address critical gaps in adaptability, scalability, and transparency for neural network-based threat detection systems. Our results highlight the potential of hybrid deep learning architectures in building more robust and dynamic cybersecurity defenses.
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4411463436