Robustness is a property of [[ Neural Network ]]s that describes how much a network’s output is changed by small modifications in the input.

Local robustness is concerned with not changing the prediction around specific inputs while [[ global robustness ]] is concerned with having a separation between regions with different labels in which the classifier must refuse to make a prediction.

Sources

• [[ Branch and Bound for Piecewise Linear Neural Network Verification ]]