Small Language Models for Phishing Website Detection: A Review of Cost, Performance, and Privacy Trade-Offs

Abstract

This review investigates Goldenits et al.'s (2025) empirical comparison of fifteen open small language models (SLMs) for phishing website detection, which aimed to evaluate whether locally hosted models can achieve the same accuracy as proprietary large language models (LLMs), without the prohibitive costs or privacy concerns. The authors test each model with a stratified sample of 1,000 labelled websites from a pool of 10,395 websites and measure the accuracy, precision, recall and F1 score of the results. The best local model, llama3.3:70b, achieves an F1 score of 0.893 and recall of 0.948, which is close to, but still lower than, the F1 scores above 0.95 achieved by the largest proprietary systems (Goldenits et al. 2025). This review restates the three research questions posed in this paper, assesses the evidence provided for each of the questions and situates the evidence in the context of the existing literature on cost-aware deployment (Irugalbandara et al. 2024; Kavya & Sumathi, 2024) and LLM-based phishing detection (Koide et al. 2024). It concludes that, besides the number of parameters, the deployability of an SLM depends on its architecture and on the reliability of its output format, which does not depend only on its classification skill.