Detecting Malicious URLs Using a Deep Learning Approach Based on Stacked Denoising Autoencoder

As the source of spamming, phishing, malware and many more such attacks, malicious URL is a chronic and complicated problem on the Internet. Machine learning approaches have taken effect and obtained high accuracy in detecting malicious URL. But the tedious process of extracting features from URL and the high dimension of feature vector makes the implementing time consuming. This paper presents a deep learning method using Stacked denoising autoencoders model to learn and detect intrinsic malicious features. We employ an SdA network to analyze URLs and extract features automatically. Then a logistic regression is implemented to detect malicious and benign URLs, which can generate detection models without a manually feature engineering. We have implemented our network model using Keras, a high-level neural networks API with a Tensor-flow backend, an open source deep learning library. 5 datasets were used and 4 other method were compared with our model. In the result, our architecture achieves an accuracy of 98.25% and a micro-averaged F1 score of 0.98, tested on a mixed dataset containing around 2 million samples.