Abstract

Phishing remains one of the most prevalent and effective forms of social engineering attacks in today’s online environment. Attackers typically impersonate trusted individuals or organizations to deceive users and gain access to sensitive information. Recent studies (e.g., Dandotiya et al., 2026) show that phishing incidents have continued to rise and remain a major cybersecurity threat. Most of these attacks are text-based, commonly delivered through email or SMS channels. Despite advances in machine learning, existing phishing detection systems still face important limitations. Prior work has identified challenges in existing phishing detection systems, including poor generalization, dataset bias, and vulnerability to adversarial manipulation (Kytidou et al., 2025). Although neural network–based approaches have improved detection performance, they are still not robust enough to handle evolving phishing strategies (Wilk-Jakubowski et al., 2025). These limitations suggest a need for more adaptable and generalizable detection methods. To address these challenges, this study explores the use of contrastive learning for phishing detection in text-based environments. Contrastive learning enables models to learn semantic representations by bringing similar phishing messages closer while separating dissimilar messages in the representation space, reducing reliance on labeled data and improving generalization. Prior research (Li et al., 2025) demonstrates that self-supervised contrastive learning frameworks show promising results in phishing detection tasks. Building on this work, this study proposes to train a neural text encoder using both phishing email and SMS datasets under a contrastive learning objective. By optimizing the encoder with a contrastive loss, the model is expected to better distinguish between phishing and legitimate messages in the representation space, leading to improved detection performance across different communication channels.

Download

Share

COinS