📝 Research Experience

A Robust Malicious Traffic Detection Framework with Low-quality Labeled Data

Role: Group leader

• Proposed a novel double-constrained similarity rule to construct a similarity topology graph and utilized the topological relationships to refine the low-quality labels.

• Implemented the algorithm and evaluated the performance on BoT-IoT and a real-world malicious traffic dataset, achieving an accuracy of 90% with 80% noise labels.

• Summarized the ideas and experimental results, wrote a paper, and accepted by ICC 2024.

A Generic Framework to Enhance Model Robustness for Intrusion Detection on Noisy Data Role: Group leader

• Proposed a generic label-noise-resistant framework for malicious traffic detection, which worked significantly better compared to state-of-the-art methods (1.69% and 1.86% better than the second-best algorithm and 20.86% and 12.61% better than the third-best algorithm).

• Implemented the algorithm of Gedss and evaluated the performance on two popular datasets, i.e., BoT-IoT and CICIDS2017. The accuracy of Gedss achieves 96.08% and 96.22% for the two datasets with 80% noise labels.

• Summarized the ideas and experimental results, wrote a paper, and submitted it to CSCWD 2024.

Model-Agnostic Generation-Enhanced Technology for Few-Shot Intrusion Detection in IoT

Role: Main developer

• Proposed a Model-Agnostic Generation-Enhanced Technology (MAGET) for few-shot intrusion detection based on GAN and MAML, which improves the accuracy of identifying few-shot attacks on two datasets by at least 2.2% and 1.5%, respectively compared with other related methods.

• Implemented the algorithm of MAGET and evaluated the performance on two popular datasets, i.e., BoT-IoT and CSE-CIC-IDS2018. The experiments show that MAGET possesses 94.3%/1.8% TPR/FPR and 99.8%/0.1% TPR/FPR in anomaly-based classification and 95.2% and 91.9% accuracy in signature-based classification, respectively.

DEML: Data-enhanced Meta-Learning Method for IoT APT Traffic Detection

Role: Main developer

Black-box Word-level Textual Adversarial Attack Based On Discrete Harris Hawks Optimization

Role: Assisted in writing thesis

Unknown traffic identification algorithm without threshold

Role: Group leader

• Responsible for drafting project applications and driving projects to successful completion.

• Proposed a threshold-free unknown traffic detection network for intrusion detection, which greatly reduces performance consumption by improving the discriminator network structure, realizing rejection of unknown classes and classification of known classes.

• Initiated ideas and implemented the algorithm, wrote a patent which has been published.