Graph Neural Network Phishing Detection via HTML Noise Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Phishing emails, particularly those utilizing HTML tags, pose challenges for machine learning tools due to imbalanced training data, rapid changes in phishing methodologies, and the complexity of handling HTML, text, and tabular data, leading to difficulties in effective detection.
Innovation Solution
An electronic device uses deep learning techniques, specifically graph neural networks trained with noise addition via Stochastic Gradient Langevin Descent, to analyze HTML tags and detect phishing attacks by generating a graph structure from these tags, enhancing robustness against small changes in training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning tools are used to detect phishing emails, then detection capability is improved, but the tools fail to keep up with rapidly changing phishing methodologies causing high temporality and label aging in training data
Solution Approach 1:
The patent implements dynamic graph neural network architectures that can adapt to changing phishing patterns. The system continuously updates graph representations of email HTML structures and retrain models with new data, allowing the detection system to evolve alongside phishing methodologies rather than relying on static training data
Solution Approach 2:
The patent applies data augmentation techniques and noise addition during training to prepare the model for variations in phishing patterns before deployment. By pre-training with augmented and noisy data, the model develops robustness to handle rapidly changing phishing methodologies without requiring immediate retraining
2Reliability
If deep learning models are trained with noisy data to improve robustness, then reliability against small changes is improved, but training time and computational resources increase
Solution Approach 1:
The patent applies noise addition and data augmentation at controlled levels during training rather than exhaustive processing. By applying partial noise augmentation and selective data augmentation techniques, the system achieves sufficient robustness without the full computational cost of processing all possible variations
Solution Approach 2:
The patent modifies training parameters such as noise level, augmentation intensity, and batch size to optimize the balance between robustness and training efficiency. By dynamically adjusting these parameters during training, the system achieves reliable robustness while controlling computational resource consumption and training time
Data Source
AI summary
A device, system, and method are provided for detecting an email phishing attack by training graph neural network to detect phishing emails based on hypertext markup language (HTML) tags and cascading style sheets (CSS) included in an email. Noise is added during the training of the graph neural network to make the trained graph neural networks more robust against small changes in the training data.