Leveraging Transfer Learning for Enhanced Cybersecurity Threat Detection: A Novel Approach For Identifying Anomalies and Attacks
Thukkani Swetha, Seshaiah Merakapudi · 2025
The presented threat detection system for cyber security integrates auto encoder model with PCA and DNN model (Auto encoder + PCA + DNN) to perform unsupervised security threat evaluation. The system implements the above two mentioned unsupervised methods for detecting security abnormalities and new threats detection in real-time network environments. Results of this model evaluation are compared against standard supervised learning models Logistic Regression, LightGBM and SGD Classifier using CICIoT23 dataset. The unsupervised hybrid model outperforms supervised models in accuracy measurements and maintains either equal or better recall and F1 score levels. The hybrid approach demonstrated remarkable performance because it achieved 0.95 accuracy and 0.85 precision together with 0.94 recall and 0.89 F1 score thus proving its effectiveness in detecting anomalies. The supervised models delivered accuracy measurements of approx. 0.87 and recall performance spanning from 0.88 to 0.89 along with the highest F1 score reaching 0.89. Findings indicate that hybrid auto encoder and PCA approaches from unsupervised learning succeed in resolving cybersecurity anomaly detection problems by demonstrating superior performance than supervised models especially within unknown attack detection and evolving attack situations. This research demonstrates how hybrid unsupervised solutions represent a promising security solution for threat detection operations in cybersecurity.