Countering AI-Driven Adversaries: An Adaptive DeceptionFramework for Resilient Web Security

This paper presents a novel deception-based cybersecurity framework that redefines web defense through adaptive, embedded traps designed to detect and contain AI-driven automated adversaries. With the rapid advancement of machine learning (ML) and large language ‎models (LLMs), traditional web security measures—such as CAPTCHA, honeypots, and Web Application Firewalls (WAFs)—have become increasingly ineffective. Modern bots can now simulate human browsing, execute JavaScript, and evade detection through the use of ‎adaptive algorithms. The proposed framework introduces invisible, dynamically generated traps within the Document Object Model (DOM) ‎and JavaScript layers of web applications to exploit the behavioral disparities between genuine users and automated systems.‎ These traps include hidden forms, off-screen anchor links, and randomized decoy endpoints that are imperceptible to human users but ‎tectable by automated bots. Interactions with these traps trigger behavioral analysis routines that calculate a Non-Human Interaction Likelihood (NHIL) score, a novel session-level metric that employs sigmoid activation functions to measure behavioral abnormality across multiple parameters. Based on this score, the system classifies, logs, and isolates non-human activity in real time.‎ An experimental evaluation in a test web environment demonstrated perfect classification performance, with an F1-score of 1.0, achieving ‎complete detection accuracy without false positives or degradation in user experience. Page load latency increased by less than five milliseconds, confirming the framework’s lightweight and seamless integration.‎ By merging adversarial design, behavioral analytics, and adaptive deception, the proposed system establishes a resilient, intelligence-driven ‎approach to web security. It transforms defensive architecture from reactive to proactive—detecting, engaging, and neutralizing automated ‎adversaries at the interaction layer—offering a scalable model for the next generation of deception-centric cybersecurity‎.