Simple but Smart: Against the Pursuit of Endless Complexity in Deep Learning Models for Detecting Phishing URLs
Musa Ibrahim Anda, Armaya’u Zango Umar* and Barira Hamisu
ABSTRACT
Phishing attacks are one of the most prevalent and evolving cyber security threats. The attacks often employ a fraudulent Universal Resource Locator (URL) and a well-crafted social engineering tactic to mislead gullible individuals into releasing their sensitive information, such as bank or credit card details. Machine learning classification models have been proposed to proactively detect phishing URLs. In an attempt to improve the detection accuracy, deep learning models were also proposed. Yet, to push the performance of the models even further, more sophisticated deep learning architectures were proposed. Nonetheless, the sophistication in the architectures does not improve the performance commensurate with its complexity. To this end, this paper compared the performance of a simple Feedforward Neural Network (FNN) against more complex architectures: hybridized Deep Neural Network and Bidirectional Long Short- Term Memory (DNN-BiLSTM), hybridized Deep Neural Network and Bidirectional Long Short-Term Memory (DNN-BiLSTM) with a transformer, and hybridized Deep Neural Network and Bidirectional Long Short-Term Memory (DNN-BiLSTM) with semantic Natural Language Processing (NLP) features. The models were trained on a phishing dataset that had label noise corrected with Cleanlab – A Confidence learning framework. The results show that a simple Feedforward Neural Network, when trained on cleaned data, can equal or even surpass the performance of any complex deep learning architecture while maintaining significantly lower runtime.
