Search for: All records

Phishing attacks remain one of the most prevalent and pervasive cybersecurity concerns. Voice phishing (i.e., vishing) is an emerging type of phishing attack where malicious actors use audio channels to steal sensitive information from victims. However, vishing detection is a challenging task due to its real-time nature and the limited availability of datasets. To help address the concern of vishing detection, this study proposes the vishing generative pretrained transformer (VishGPT). VishGPT adopts the computational design paradigm and incorporates novel reinforcement learning-based large language model fine-tuning and synthetic data model pretraining to automatically detect vishing attempts in real time. We evaluated VishGPT using a series of benchmark experiments, where we empirically demonstrated its improvement over state-of-the-art vishing detection and audio classification models. The results suggest that our proposed VishGPT achieved state-of-the-art performance in terms of accuracy (86.18%), precision (90.63%), recall (85.02%), and F1-score (87.74%). VishGPT offers practical value to cybersecurity professionals, end users, and academia. Additionally, VishGPT provides important design principles in the form of a custom proximal policy optimization (PPO) reward function and synthetic pretraining to the information systems knowledge base. 

Artificial intelligence (AI) increasingly powers sensitive applications in domains such as healthcare and finance, relying on both extit{linear operations} (e.g., matrix multiplications in large language models) and extit{non-linear operations} (e.g., sorting in retrieval-augmented generation). Fully homomorphic encryption (FHE) has emerged as a promising tool for privacy-preserving computation, but it remains unclear whether existing methods can support the full spectrum of AI workloads that combine these operations. In this SoK, we ask: extit{Can FHE support general AI computation?} We provide both a functional analysis and a cost analysis. First, we categorize ten distinct FHE approaches and evaluate their ability to support general computation. We then identify three promising candidates and benchmark workloads that mix linear and non-linear operations across different bit lengths and SIMD parallelization settings. Finally, we evaluate five real-world, privacy-sensitive AI applications that instantiate these workloads. Our results quantify the costs of achieving general computation in FHE and offer practical guidance on selecting FHE methods that best fit specific AI application requirements. Our codes are available at https://github.com/UCF-ML-Research/FHE-AI-Generality. 

Protected Health Information (PHI, e.g., electronic health records, insurance information) is increasingly stolen in data breaches by malicious actors with the intent to sell to others in hacker communities. These actors often protect themselves by describing the content and availability of PHI data using encrypted messaging platforms (e.g., Telegram & Discord). However, the extent and nature of these PHI discussions are not well known. Therefore, in this research, we propose a Named Entity Recognition Framework for PHI (NERF-PHI) to systematically analyze PHI-related hacker conversations. To conduct our research, we collected more than three million multilingual hacker posts from Discord servers and Telegram groups. Utilizing open-source machine translation tools, we translated conversations to English and extracted information related to vulnerable individuals and medical entities. Results from our study suggest that encoder-based Large Language Models show significant promise for extracting PHI-related information from hacker communities and can be used by cybersecurity professionals and law enforcement to combat PHI misuse. Our study is also one of the first comprehensive analyses of multilingual PHI discussions in hacker communities.