Abstract

Phishing attacks remain one of the most prevalent cybersecurity threats, exploiting users through deceptive techniques to obtain sensitive information. Existing deep learning approaches such as ResNeXt-GRU have demonstrated high detection accuracy but suffer from computational overhead, limiting their applicability in real-time environments. In addition, their reliance on static hyperparameter optimization techniques limits their ability to adapt to emerging phishing attack patterns, thereby reducing their effectiveness in dynamic real-time environments. In this study, a lightweight deep learning framework for phishing detection is proposed. The model introduces a modified ResNeXt-GRU architecture with pruning and quantization techniques to reduce computational complexity while maintaining accuracy. The proposed framework utilizes URL-based features and deep feature extraction for effective classification. The framework was evaluated through simulation using the PyTorch 2.0+ core libraries with TensorFlow backends to assess its performance. Performance evaluation was conducted using standard metrics, including accuracy, precision, recall, F1-score, and execution time. Experimental results show that the proposed model achieved reduced computational cost with higher accuracy, making it suitable for deployment in resource-constrained environments.

Key words: Phishing Detection, Deep Learning, ResNeXt-GRU, Pruning, Quantization