D ∈ {X, y}: dataset

f1: model to enhance

f2: model to avoid enhancement

nfolds: number of folds for K-fold partitioning

λ: enhancement step size

η: enhancement suppression coefficient

for k = 1 : nfolds do

Establish Dtrain , Dheld-out

Train f1, f2

g1 ← ∇xA(f1,Xheld-out)

g2 ← ∇xA(f2,Xheld-out)

$X_{\mathit{\text{held-out}}}\leftarrow X_{\mathit{\text{held-out}}}-\lambda \left({\mathit{\text{proj}}}_{g_2^{\perp }}\left(g_1\right)-\eta {\mathit{\text{proj}}}_{g_1^{\perp }}\left(g_2\right)\right)$

end for