New Optimized Adaptive Time Series IDS Classifier Algorithm: Beyond Deep Learning

Adaptive time series Intrusion Detection System (IDS) Classifier is essential to detect real-time cyber-threats. Meanwhile, optimized hyperparameters on time series IDS classifier model will ensure swift detection. However, current studies on Time Series IDS classifier involve additional RNN-LSTM layers and multiple gates to optimize the training and feedback process. Notwithstanding, RNN-LSTM has powerful features to memorize data sequences. The nature of multiple complex hidden states in RNN-type model requires intensive training or epoch to achieve optimized loss function. This paper aims to go beyond conventional deep learning model by removing complex gated states and conventional hidden layers. The goal is to create an optimized adaptive time series classifier. The model leverages various fitting algorithms which include Sinusoidal, Linear, Power Function, Taylor Series and a new “Staircase” function that is introduced in this study. These functions adapt gradually to the real-time target distribution pattern. This will eliminate the need for feedback process to optimize hyperparameters. The model’s performance is evaluated against the realistic benchmarked IDS dataset; a dataset that simulates recent malware attacks and has imbalanced distribution property. This property reflects a realistic low cyber-attack footprint. After 10 epoch over randomized stratified testing samples, the Mean Absolute Error (MAE) rate achieved almost 0.0% after a fitting process reached 100% as compared with the conventional LSTM model that achieved 17%.