Using multimodal model consistency to detect adversarial attacks

This invention describes a way to protect artificial intelligence models from being tricked by malicious inputs, known as adversarial attacks. It works by using data that exists in multiple forms, such as a picture and a text description of that picture. If an attacker tries to subtly alter one form of the data (like the image) to fool the AI, this system checks if the altered data still makes sense when compared to the other, unaltered form. If the difference between the two forms is too great, indicating an inconsistency, the system rejects the suspicious input, preventing the AI from making a mistake.