Algorithmic deception: a comparative entropy-based analysis of GenAI-driven phishing and human social engineering
Nature Journal of Emerging Sciences Technologies and Innovations · 28.05.2026 · 8 (4) · 377-390
DOI: 10.65752/9173v732 ↗Kurz gefasst
Fachartikel in der Zeitschrift Nature Journal of Emerging Sciences Technologies and Innovations, online erschienen am 28.05.2026. Verfasst von Abdullahi Kabir Muhammed und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The development of large language models (LLMs) has transformed the threat landscape of social engineering by enabling attackers to generate highly fluent, context-aware phishing emails at scale. As grammatical errors and stylistic inconsistencies diminish, traditional linguistic detection cues become unreliable. This study reframes AI-driven phishing detection as a problem of algorithmic entropy, investigating whether content-agnostic stylometric signals can distinguish LLM-generated emails from human-authored communication without inspecting semantic content. Using a balanced dataset of 4,000 human and LLM-generated emails across phishing and legitimate categories, we extract Shannon-entropy–derived proxy features, including GPT-2 perplexity, sentence-length burstiness, type–token ratio, and readability metrics. An interpretable logistic regression classifier is trained on a length-matched, deduplicated evaluation split to eliminate structural artifacts. Results demonstrate clear separability in entropy space, achieving a ROC-AUC of 0.980 and 95.1% accuracy on a held-out test set, with comparable robustness under phishing-only evaluation (ROC-AUC = 0.959), while a length-only baseline performs near chance (ROC-AUC ≈ 0.61), confirming that discrimination arises from intrinsic statistical regularities rather than superficial artifacts. These findings provide empirical support for entropy-based stylometry as a lightweight, privacy-preserving attribution mechanism capable of complementing existing email security systems. The study further discusses robustness considerations and implications for adversarial adaptation in evolving AI-enabled phishing campaigns.
Zitieren
APA
Muhammed, A. K., Ifeanyi, F., Adamu-Fika, F., & Oluwaseyi, E. O. (2026). Algorithmic deception: a comparative entropy-based analysis of GenAI-driven phishing and human social engineering. Nature Journal of Emerging Sciences Technologies and Innovations, 8(4), 377-390. https://doi.org/10.65752/9173v732
Vancouver
Muhammed AK, Ifeanyi F, Adamu-Fika F, Oluwaseyi EO. Algorithmic deception: a comparative entropy-based analysis of GenAI-driven phishing and human social engineering. Nature Journal of Emerging Sciences Technologies and Innovations. 2026;8(4):377-390. doi:10.65752/9173v732
BibTeX
@article{Muhammed2026,
author = {Muhammed, Abdullahi Kabir and Ifeanyi, Fortune and Adamu-Fika, Fatimah and Oluwaseyi, Ezekiel Olorunshola},
title = {Algorithmic deception: a comparative entropy-based analysis of GenAI-driven phishing and human social engineering},
journal = {Nature Journal of Emerging Sciences Technologies and Innovations},
year = {2026},
volume = {8},
number = {4},
pages = {377--390},
doi = {10.65752/9173v732}
}
EUVIMED