The performance loss a model experiences when trained to be robust against attacks instead of optimized purely for accuracy.