An Ensemble of Activation Functions in AutoEncoder Applied to IoT Anomaly Detection

We propose an ensemble of activation functions for learning the latent representation of AutoEncoder (AE) to improve the accuracy of Internet of Things (IoT) botnet (anomaly) detection. The proposed model is called Ensemble Latent Representation (ELR) that combines the activation functions, i.e., hyperbolic Tangent (Tanh) and Rectified Linear Unit (Relu) in hidden layers of AE. This model enhances the advantages and mitigates the shortcomings of Tanh and Relu function. As a result, the latent representation learnt by AE is more robust, thereby improving the accuracy of classification algorithms for the IoT anomaly detection. To evaluate the effectiveness of the proposed model, we carried out intensive experiments on four IoT anomaly datasets using four popular classifiers, i.e., Support Vector Machine (SVM), Perceptron (PCT), Nearest Centroid (NCT), and Linear Regression (LR). The experimental results show that our proposed model significantly improves the accuracy of IoT anomaly detection up to 19.9% compared to the original one using area under the curve metric.