The rapid growth of online services has heightened concerns about user protection from cyber threats, particularly phishing, which poses significant risks to cyber-social security. To this end, we propose a novel tool for phishing detection called U-Proof. Our tool uses both state-of-the-art LLMs and traditional ML models to detect phishing websites. In particular, we evaluate the phishing detection capabilities of different LLMs and compare them with several ML models to analyze the impact of different model architectures on the identification of phishing websites. For a comprehensive experimental evaluation, we use a combination of public and custom datasets. These include active phishing websites from September 2024, as well as URLs from banks and postal services. Furthermore, the tool includes explanations to enhance user awareness of phishing tactics, supporting broader educational efforts to reduce risks.

Recently, dishonest sellers are using social platforms to advertise products that can be purchased for free through a refund mechanism, which is based on the writing of five-star fake reviews. The aim is to increase product visibility by influencing their ranking compared to similar products. This mechanism is leading to a significant distortion of e-commerce platforms, eroding trust among customers and sellers. In this paper, we address the problem of identifying fake reviews, aiming to provide an approach for mitigating fraudulent practices that compromise the integrity and transparency of e-commerce platforms. We propose a supervised model tailored for identifying fake reviews for refund purposes and compare its performance with some of the most recent generative models. Since, to the best of our knowledge, no datasets exist in the literature suitable for fake review identification in the process of Purchasing, Requesting reviews, and Refunding a product, we first proposed a new dataset of fake and genuine reviews from Amazon, collected with the help of a domain expert. Then, we defined five other new datasets containing reviews automatically generated by language models. To interact with these models, we designed new prompt approaches specifically tailored to our goal, which exploit the iterative refinement behind these models for improving classification results. Experimental results demonstrated the effectiveness of the supervised model in detecting both types of fake reviews, outperforming state-of-the-art models with improvements ranging from 0.23 to 0.70 in terms of accuracy, precision, and recall. • The requesting reviews and refunding phenomenon (PRP process) is analyzed. • PRP reviews are collected from web pages and then validated by a domain expert. • A BERT-based model is successfully applied for properly identifying PRP fake reviews. • The proposed model is compared with Large Language Models and state-of-the-art models. • New datasets of real and automatically generated PRR reviews have been proposed.

Intelligent diagnosis processes rely on Artificial Intelligence (AI) techniques to provide possible diagnoses by analyzing patient data and medical information. To make accurate and quick diagnoses, it is possible to use AI tools to efficiently analyze huge amounts of data and find patterns that a clinician might miss. In recent years, new large language models (LLMs), such as ChatGPT and Google BARD, have shown remarkable capabilities in several domains, including intelligent diagnostics. This research aims to compare the performances of ChatGPT and traditional machine learning models for making diagnoses of low- and medium- risk diseases only based on their symptoms. On the basis of our study, we defined four research questions: RQ1) What are the benefits and limitations of using ChatGPT in intelligent diagnosis? RQ2) How do traditional machine learning approaches compare to ChatGPT for intelligent diagnosis? RQ3) How does ChatGPT compare with other LLMs and domain-specific natural language processing models in the intelligent diagnosis tasks?, and RQ4) What are the implications of the predictive models and ChatGPT for healthcare, and how can they be used to support people?. To answer these RQs, we first evaluate the performances of different engines of ChatGPT, also introducing a new prompt engineering methodology specifically tailored for achieving accurate diagnostic outcomes. Moreover, we compare these results with those achieved by different predictive models trained for intelligent diagnosis tasks, i.e., Google BARD, and two domain-specific NLP models. Finally, we propose a new interactive bot available for users that relies on the best-performing models evaluated in the previous steps. The experiments have been conducted using two medical datasets for disease prediction consisting of more than 100 symptoms associated with several diagnoses.