Federated and Transfer Learning-Based Network Intrusion Detection Systems (NIDS) for Cyber Attack Detection
This project addresses the growing need for intelligent and privacy-aware solutions in cybersecurity by implementing Federated Learning and Transfer Learning for Network Intrusion Detection using the UNSW-NB15 dataset. The objective was to build a system capable of accurately detecting various cyber-attacks while overcoming the limitations of traditional machine learning methods.
Traditional machine learning techniques, while effective in controlled environments, require large centralized datasets, struggle with imbalanced data, and do not support distributed or privacy-sensitive deployments. These constraints make them less suitable for modern, real-world cybersecurity applications.
To overcome these limitations, we implemented the following approaches:
Transfer Learning: Leveraged pre-trained models (XGBoost, Autoencoders, LSTM) to improve detection accuracy in scenarios with limited or imbalanced labeled data. Transfer learning reduced training time and boosted generalization.
Federated Learning: Enabled training across decentralized data sources without transferring raw data. Using algorithms like Random Forest, Decision Trees, and CNN in a federated setup, we ensured privacy while maintaining strong model performance in distributed environments.
Challenge 1: Data Imbalance
Solution: Applied techniques like SMOTE and class weighting to improve model performance on minority attack classes.
Challenge 2: Data Privacy and Distribution
Solution: Used Federated Learning to train models collaboratively without sharing raw data across devices or institutions.
Challenge 3: Model Convergence and Resource Constraints
Solution: Selected efficient models like Random Forest and XGBoost, and fine-tuned federated settings to balance accuracy with computational efficiency.
Improved Generalization: Transfer learning significantly outperformed traditional models trained from scratch by reusing learned features from related domains.
Privacy-Preserving: Federated learning allowed training on sensitive data locally, something that normal ML models cannot do.
Scalable and Real-World Ready: Our methods aligned with modern, distributed cybersecurity infrastructures, unlike traditional centralized ML.
Efficient on Imbalanced Data: Traditional models are often biased toward majority classes, while our approach enhanced detection across all attack types.
CONCLUSION :
This project successfully demonstrated the application of Federated Learning and Transfer Learning to develop a robust, scalable, and privacy-preserving Network Intrusion Detection System (NIDS).The findings pave the way for future research in deploying advanced machine learning models in distributed, real-time, and privacy-constrained cybersecurity systems.
At Lendi, we are recognized as a top institution dedicated to excellence in education.we offer exceptional academic programs that prepare our students for global success. Discover the unique qualities that make our students stand out.
©Lendi Institute of Engineering and Technology - Established 2008