Adaptive Anomaly Detection Using Deep Bi-LSTM and Self-Adaptive Pufferfish Optimization

The task of anomaly detection has become critical in many applications, including network security, fraud detection, and fault diagnosis. Traditionally, most of the existing methods fail with high-dimensional data, making it difficult and challenging to get accurate anomaly detection with low false positives. The main techniques proposed so far may also face problems in being computationally costly and not adaptive to the variability of the data characteristics. Therefore, there is an urgent need to design a well-built, efficient, and adaptive framework that properly identifies anomalies within large volumes of complex datasets while optimizing the selection of features. To solve these issues, this paper proposes a novel advanced preprocessing, feature extraction, feature selection, and classification methods integrated into the anomaly detection framework. Initially, the dataset undergoes data cleaning, normalization, and standardization to ensure uniformity in scale and distribution. Then it will extract statistical features that encapsulate critical insights from the raw data. For the detection stage, we propose Deep BiLstm, a hybrid deep learning architecture that leverages the strengths of Bi-Lstm by using lightweight densely connected layers for efficient anomaly classification. The hyperparameter of the Deep Bi-Lstm model is fine-tuned using the Self-Adaptive Pufferfish Optimization Algorithm (SAPFOA). The entire framework is implemented using the Python tool. Evaluate the effectiveness of our approach over the Ton-IoT dataset, which is the most frequently used benchmark of network intrusion and anomaly detection.