Randomized Smoothing in AI Explainer
Randomized smoothing turns a base classifier into a smoothed classifier by adding random noise to an input many times. The final prediction is the class that the base model chooses most often across those noisy copies.
If one class wins by a sufficiently large margin, statistical bounds can certify that small changes within a particular distance cannot alter the smoothed prediction. Gaussian noise is commonly used for guarantees under an L2 norm.
The method can scale to models that are too large for some exact verification techniques. It also introduces extra computation because reliable estimates require many noisy evaluations, especially when probabilities are close.
The certificate depends on the noise distribution, sampling confidence, and threat model. More noise may increase the certified radius while reducing ordinary accuracy, so the operating point is a deliberate tradeoff.
Randomized smoothing builds a more robust classifier by averaging a base model's predictions across many noisy versions of the same input.
Key terms include base classifier, smoothed classifier, Gaussian noise, vote, confidence bound, certified radius, L2 norm, sampling and tradeoff.
The same input is evaluated repeatedly after random noise is added.
The most probable class becomes the smoothed classifier's answer.
Stronger smoothing can protect a wider region but blur useful signal.
Sampling cost and statistical confidence also affect practical deployment.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.