Human in the loop reinforcement learning framework for adaptive phishing detection

Abstract

Phishing attacks remain a persistent threat in cybersecurity, exploiting human psychology to extract confidential information. Conventional machine learning and Reinforcement Learning (RL) approaches often rely on rigid binary reward structures that oversimplify phishing behavior and demand extensive datasets, limiting their practical deployment. This paper presents HARE (Human-guided Adaptive Reinforcement for Engineering), an RL-based framework that integrates human feedback to adaptively shape the agent’s reward signal. By incorporating scalar, context-aware guidance from experts, HARE enables faster convergence with fewer samples, improves adaptability to evolving phishing tactics, and captures subtle behavioral cues frequently missed by automated detectors. Experimental evaluation on the publicly available PhiUSIIL dataset shows that HARE achieves an F1-score of 97.62% with a false positive rate of 2.55%, confirming its reliability and data efficiency for adaptive phishing detection.