MEMCAIN: a memory-enhanced hybrid CNN-attention model for network anomaly detection

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Network Security

Background:

• Increasing cybersecurity threats necessitate advanced network defense mechanisms.
• Current deep learning models for intrusion detection face challenges like class imbalance and limited feature representation.
• Single-task frameworks hinder the potential for multi-task collaboration in deep learning for cybersecurity.

Purpose of the Study:

• To propose MEMCAIN, a multi-task feature fusion deep learning method to overcome limitations in current intrusion detection systems.
• To enhance the accuracy and robustness of intrusion detection by addressing class imbalance and improving feature extraction.
• To leverage multi-task learning for synergistic performance gains in network security.

Main Methods:

• Developed MEMCAIN, integrating Convolutional Neural Networks (CNN) with attention mechanisms (CCA Blocks) for spatiotemporal feature extraction.
• Employed a memory autoencoder to capture latent distribution features of network traffic flows.
• Implemented an end-to-end collaborative training framework to jointly optimize the CNN-attention network and the memory autoencoder.

Main Results:

• MEMCAIN demonstrated significant superiority over existing baseline methods across multiple datasets.
• Ablation studies confirmed the efficacy of individual modules within MEMCAIN for fine-grained intrusion detection.
• The proposed method effectively mitigates issues of class imbalance and enhances feature representation.

Conclusions:

• MEMCAIN offers a powerful and effective solution for fine-grained intrusion detection in complex network environments.
• The multi-task feature fusion approach significantly improves detection performance compared to single-task methods.
• This research contributes to advancing deep learning applications in cybersecurity for more resilient network protection.