BEYOND ACCURACY

Beyond Accuracy examines one of the most important challenges in AI-based phishing email detection: a model may perform well in controlled experiments, yet fail when exposed to new datasets, real-world variation, adversarial manipulation, multilingual content, or changing attack strategies. This book argues that high accuracy alone is not enough. What matters is whether a system remains reliable, robust, and trustworthy when conditions change. The book introduces TRACE-PHISH, an original framework for evaluating the robustness and generalization of artificial intelligence models for text-based phishing detection. It brings together adversarial testing, cross-dataset validation, multilingual evaluation, explainability, and generative AI-based stress testing into a single structured approach. Rather than asking only whether a model works, the book asks a more practical and important question: under what conditions can its performance be trusted? Drawing on a critical analysis of the scientific literature and current developments in AI and cybersecurity, the book discusses machine learning models, transformer-based systems, and large language models used for phishing detection. It shows the main weaknesses of current evaluation practices, including overreliance on benchmark datasets, limited external validation, insufficient robustness testing, and weak attention to operational deployment risks. This work is intended for researchers, postgraduate students, cybersecurity professionals, and decision-makers interested in artificial intelligence, phishing defence, and trustworthy security evaluation. By moving the discussion from simple benchmark success to real-world assurance, Beyond Accuracy offers both a conceptual contribution and a practical roadmap for building more dependable phishing detection systems.