A Multilingual Behavioral Speech Dataset for Vishing and Social Engineering Detection

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Speech Technology

Background:

• Voice phishing (vishing) is a significant social engineering threat, leveraging speech's persuasive elements.
• Existing datasets for vishing research are often limited, being monolingual and lacking detailed annotations for telephony interactions.
• There is a need for robust, multilingual resources to study and detect persuasion in vishing calls.

Purpose of the Study:

• To introduce VISHGUARD, a novel multilingual synthetic audio corpus for advancing research in persuasion-sensitive vishing detection.
• To provide a structured dataset with multidimensional annotations to facilitate controlled studies on vishing tactics.
• To support the development of more effective vishing detection systems in diverse linguistic contexts.

Main Methods:

• Generation of 3,000 simulated phone calls using text-to-speech synthesis and controlled noise mixing.
• Inclusion of English, French, and Modern Standard Arabic, with balanced distribution between fraudulent and legitimate calls.
• Annotation of each audio sample for persuasion strategy, interactional markers, and emotional tone, with durations ranging from 10 to 90 seconds.

Main Results:

• The VISHGUARD corpus offers a reproducible, script-conditioned synthetic telephony dataset.
• It features balanced linguistic and class distributions, crucial for reliable model training and evaluation.
• Multidimensional annotations enable fine-grained analysis of persuasive elements in simulated vishing calls.

Conclusions:

• VISHGUARD addresses the limitations of existing datasets, providing a valuable resource for multilingual vishing research.
• The corpus facilitates the development of persuasion-aware vishing detection models.
• Public availability ensures transparency and promotes further research and innovation in cybersecurity.