Anomaly detection of cybersecurity behavior using cross-sequence aligned transformer-A dynamic recognition approach...

Area of Science:

• Cybersecurity
• Network Security
• Artificial Intelligence

Background:

• High-frequency network environments present challenges for anomaly detection due to temporal asynchrony and data redundancy in traffic and user behavior.
• Existing models struggle to identify dynamic attack patterns effectively in these complex scenarios.

Purpose of the Study:

• To propose and validate a novel deep learning model for enhanced cybersecurity anomaly detection in high-frequency interaction networks.
• To address the limitations of current methods in handling temporal asynchrony and information redundancy.

Main Methods:

• Developed a Cross-Sequence Aligned Transformer-driven Dynamic Recognition Model (CSAT-DRM).
• Employed a cross-sequence alignment mechanism to correlate multi-source time-series data without losing temporal information.
• Integrated an interaction-sensitive residual structure and dynamic thresholding for improved anomaly discrimination.

Main Results:

• CSAT-DRM achieved high performance metrics: 0.968 accuracy, 0.957 precision, 0.953 recall, and 0.955 F1-score.
• The model significantly outperformed baseline methods like LSTM, CNNs, Transformer, and CNN-BiLSTM.
• Demonstrated effectiveness in detecting both burst and persistent anomalies with high stability.

Conclusions:

• Explicitly modeling temporal asynchrony and performing collaborative modeling on a unified temporal scale enhances anomaly detection accuracy and stability.
• The CSAT-DRM offers a feasible and generalizable technical pathway for real-time threat identification in complex network environments.