Adversarial Attacks are best described as?

• A. Specific Attacks Designed To Trick The AI System, Such As Misinterpreting A Stop Sign
• B. Random Data Corruption
• C. Hardware Faults Causing Model Crashes
• D. Honest Mistakes In Labeling Data

Answer: (A)

Adversarial Attacks exploit weaknesses in AI models by feeding inputs that are intentionally designed to cause the system to make a wrong decision, often while the input looks harmless to humans. For example, a stop sign image can be subtly altered with carefully calculated changes that humans barely notice, yet cause a vision model to misclassify it as a different sign. These perturbations are purposeful, aimed at steering the model’s output toward a specific incorrect result or any incorrect result.

This is different from random data corruption, hardware faults, or labeling mistakes, which are accidental and not crafted to manipulate the model’s decisions. So the description that emphasizes deliberately crafted inputs meant to trick the AI system best captures what adversarial attacks are.