Multimodal zero-shot classification for AI applications in cybersecurity
Klaus Schwarz, Reiner Manfred Creutzburg, Franziska Schwarz, Kendrick Bollens · 2025
In this paper, we explore the application of zero-shot text classification for cybersecurity, focusing on its ability to detect threats without requiring labeled training data. Traditional cybersecurity systems rely on pre-identified threats and large, labeled datasets to function effectively. However, the evolving nature of cyber threats, such as phishing attacks and data breaches, often leaves security teams struggling to keep up. Our approach leverages pre-trained language models, allowing for real-time classification of potentially harmful content in various communication formats, such as emails, social media, and reports, without needing specific task-related data. This flexibility makes zero-shot learning a valuable tool for identifying new and emerging threats, enhancing cybersecurity infrastructure with minimal setup and maintenance. We demonstrate the practicality of this method through various use cases, showing how it can quickly categorize potential risks in a range of contexts. The results highlight the effectiveness of zero-shot classification in reducing both time and resources needed for threat detection while improving overall security outcomes. Our approach is particularly well-suited for environments where labeled data is scarce, offering a scalable and adaptable solution for modern cybersecurity challenges.