L-MACD: Learning-Based Adaptive CAPTCHA Defense Against Multimodal AI Attacks

The high rate of development of multimodal AI models such as GPT-4V, Gemini 1.5, and Claude 3 has rendered the conventional CAPTCHA schemes far less efficient. In order to address this challenge, the paper introduces L-MACD (Learning-based Multi-modal Adaptive CAPTCHA Defense) a reinforcement learning architecture which modifies the difficulty of CAPTCHA in real-time. The system uses feedback as input continuously to modify its complexity to adapt to the specific user input through the simulated human user and also the AI agents to adjust its the most important parameters like visual distortion, background noise or image rotation. It predicts the most effective defense configuration to be applied in every situation using an adaptive model implemented as a random forest algorithms. Also, a new metric has been proposed, namely Adversarial Gain Equilibrium (AGE), to quantify the extent to which a system remains human-usable and resistant to automated attacks. Four configurations of the framework were tried, being Baseline (G1), LLM-ZSA-CG (G2), Human-Expert (G3) and RL-Adaptive (G4).The findings indicate that L-MACD decreases the AI success rate from 68% to 8% (approximately 85% reduction), while maintaining 85–90% human usability. A t-test supported approach to statistical analysis shows that RL-Adaptive approach is better than static base methodologies. Additional research with SHAP-based explainability shows that rotation and distortion parameters have the greatest effect on adversarial gain reduction. To summarize, L-MACD is a versatile and transparent defense mechanism, which is capable of combating the increasing threat of sophisticated multimodal AI systems.