Adversarial inputs are carefully designed changes to data that confuse AI models and cause them to make mistakes. These changes can be subtle and hard for humans to detect, but they exploit weaknesses in the AI’s understanding. Adversarial inputs can reduce the accuracy and reliability of AI, so defending against them is important to ensure AI makes correct and safe decisions.

‍