NeuroSymbolicPhishDefend for adaptive multimodal phishing detection against evolving artificial intelligence driven...

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Machine Learning

Background:

• Phishing attacks are increasingly sophisticated due to AI and automation, making traditional detection methods insufficient.
• Current deep learning models struggle with adversarial attacks and lack interpretability for security analysts.

Purpose of the Study:

• To develop a robust and interpretable phishing detection system capable of handling advanced, AI-generated threats.
• To improve the generalization and resilience of phishing detection models against evolving attack vectors.

Main Methods:

• A neuro-symbolic multimodal framework integrating textual, visual, and metadata features using cross-attention fusion.
• Incorporation of symbolic reasoning to enhance decision-making consistency for obfuscated and adversarial phishing patterns.
• Utilized diffusion-based adversarial augmentation and an online adaptation module for continuous learning and resilience against AI-generated content.

Main Results:

• Achieved up to 97% ROC-AUC on clean test data, demonstrating high detection accuracy.
• Showcased enhanced resilience to adversarial perturbations and 6-7% absolute AUC gains in cross-dataset generalization compared to baselines.
• The SHAP-based explainability module provided transparent and interpretable feature-based predictions.

Conclusions:

• The proposed neuro-symbolic multimodal system offers a practical and reliable solution for advanced phishing detection.
• The framework's robustness, generalization, and interpretability make it suitable for integration into security infrastructures like email filters and web browsers.