What Are Adversarial Examples?
What you'll learn: How tiny, carefully crafted changes to inputs can cause models to make confident but completely wrong predictions.
The Core Concept
Adversarial examples are inputs that have been deliberately modified with imperceptible perturbations—changes so small that humans can't detect them—yet these tiny alterations cause machine learning models to misclassify with high confidence.
Imagine a stop sign. To you, it's clearly a stop sign. But add a few strategically placed stickers that look like random noise, and a self-driving car's vision system might confidently classify it as a speed limit sign. The perturbation is invisible or meaningless to humans, but catastrophic to the model.
Why They Matter
These examples expose a fundamental vulnerability in deep learning models. Unlike human perception, which is robust to small changes, neural networks can be extremely sensitive to precise pixel-level modifications. The model doesn't just make a mistake—it makes it with high confidence, which makes the failure particularly dangerous.
Key Characteristics
- Imperceptibility: Changes are typically bounded by a small ε (epsilon) value, making them undetectable to humans
- Transferability: Adversarial examples crafted for one model often fool other models too
- Confidence: Models assign high probability scores to incorrect classes
- Intentionality: Unlike random noise, adversarial perturbations are specifically optimized to cause misclassification
This connects directly to why adversarial robustness matters—models deployed in security-critical applications (autonomous vehicles, medical diagnosis, authentication) must resist such attacks.
Key Takeaway: Adversarial examples are imperceptible input modifications that exploit model vulnerabilities to cause confident misclassifications, revealing that deep learning models perceive data very differently from humans.