Zero-day attack detection in multivariate time series is an increasingly vital field, driven by its significance in security-sensitive environments, such as network monitoring. Traditional anomaly detection methods often depend on predefined patterns, which are ineffective against zero-day attacks that exploit previously unidentified vulnerabilities. To address this limitation, we introduce ZAD-ML, an unsupervised learning framework designed specifically for detecting zero-day attacks by utilizing behavioral analytics without prior knowledge of attack signatures. ZAD-ML incorporates a dual-layer neural network system where the first layer learns to compress and encode normal temporal behavioral patterns into a dense representation, facilitating efficient anomaly detection and allowing it to adapt and update its understanding of normal behavior continuously. The second layer is enhanced with attention mechanisms to analyze temporal sequences for behavioral deviations, allowing the system to adaptively update its baseline for normal behavior based on emerging data trends. We incorporate deep learning techniques to enhance the model’s ability to learn complex patterns and anomalies in data behavior. We evaluated our framework on four public Network datasets, demonstrating its capability to detect zero-day attacks with high accuracy and significantly reduced false positives as compared with existing methods. ZAD-ML provides a robust, adaptable solution for real-time anomaly detection. The implementation of our proposed method is now publicly accessible at https://github.com/don2c/ZAD-ML.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号