Phishing detection models vulnerable to adversarial attacks, study finds

A new study published on arXiv compares the effectiveness of two machine learning models, TF-IDF + Logistic Regression and a fine-tuned DistilBERT transformer, in detecting phishing emails. While both models achieved over 98% accuracy on standard test data, their performance significantly degraded when subjected to adversarial attacks designed to evade detection. The TF-IDF + LR model dropped to 64.00% accuracy, and DistilBERT fell to 63.64%, highlighting a similar vulnerability to adversarial manipulation. The research emphasizes the need for adversarial testing to be a standard part of phishing detection evaluation, as clean-data accuracy does not predict robustness against such attacks. AI

IMPACT Highlights the need for robust evaluation of AI models against adversarial attacks, particularly in cybersecurity applications.

RANK_REASON The cluster contains an academic paper detailing a comparative study of machine learning models for a specific task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• DistilBERT
• Logistic Regression
• Shap
• TF-IDF

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →