Vigilis: Leveraging Language Models for Fraud Detection in Mobile Communications and Financial Transactions

Abstract

Although advances in security have strengthened defenses in digital financial systems, attackers increasingly rely on social engineering to achieve their goals. These attacks are difficult to detect and prevent with existing security measures. To address this, we propose Vigilis, a fraud-protected application that employs advanced language models to counter such attacks in calls, texts, and payments. We first collect and make available a corpus of fraudulent calls from the Internet and train lightweight transformer-based models that achieve fraud detection accuracies of up to 94% and 87% on transcript and audio modalities, respectively. We integrate these models into a real-time call system within Vigilis that operates entirely on-device, enabling accurate fraud detection in an efficient and privacy-preserving manner. We then extend Vigilis to incorporate context-aware transaction authentication, where the underlying social context behind a transaction is determined from calls, texts, and browsing history and used to infer the transaction’s validity. By uniquely incorporating social concepts into traditional cybersecurity techniques, we attempt to counter and mitigate issues related to social engineering attacks in financial fraud.