aaebdb8bb6b0e73f6c3c54a0ab0c6415-Supplemental.pdf

Wewillshowthat there exists a classifier inH achieving0 training error on the corrupted training setSclean Sadv. We then invoke the result from Corollary4. Recall that there exists a universal constantC0 for which η d is C0-subgaussian ([18]). From this, we have that ATv γ with probability1 δ/2, which implies thath (x+η) = h (x) with probability1 δ/2 over the draws ofη. Next,recall that the following holdsforbw obtained from the minimization inthe theorem statement and for a trainingsetS Dm (see,forinstance,Theorem26.12of[5]):