How do CNNs behave under adversarial perturbations and why are they vulnerable?

CNNs are vulnerable to small adversarial perturbations because they rely on high-dimensional linear decision boundaries.

CNNs can be fooled by tiny, carefully crafted noise because their decision boundaries in high-dimensional space are often linear and sensitive. Adversarial examples exploit this by shifting inputs just enough to cross decision boundaries. Techniques like adversarial training, regularization, and input preprocessing are used to improve robustness.