Universal Adversarial Perturbations
What you'll learn: How a single carefully crafted noise pattern can fool a model on most inputs, unlike input-specific adversarial attacks.
The Core Idea
In earlier lessons, you learned about adversarial examples—small perturbations tailored to fool a model on one specific input. Universal Adversarial Perturbations (UAPs) take this further: they're single perturbations that work across many different inputs.
Think of it like this: instead of creating a custom disguise for each person to fool a face recognition system, you find one magical hat that makes almost anyone unrecognizable when they wear it.
How UAPs Differ
Input-specific attacks (like FGSM or PGD):
- Compute a unique perturbation for each input image
- High success rate on that specific input
- Don't generalize to other inputs
Universal perturbations:
- One perturbation works on ~80-90% of inputs from a dataset
- Created by iteratively testing and refining across many examples
- Typically image-agnostic: add the same noise pattern to any input
Why This Matters
UAPs reveal a systematic vulnerability in neural networks rather than input-specific weaknesses. They suggest deep learning models have shared blind spots that attackers can exploit at scale. This is particularly concerning because:
- The attacker needs to compute the perturbation only once
- It can be applied as a "universal attack vector" to many inputs
- It transfers well between similar models due to the transferability property you've seen
Key Takeaway: Universal Adversarial Perturbations are input-agnostic noise patterns that successfully fool a model on most samples from a distribution, revealing fundamental vulnerabilities in neural network decision boundaries rather than input-specific weaknesses.