Adversarial-Aware Deep Learning System Based on a Secondary Classical Machine Learning Verification Approach

. 2023 Jul 11;23(14):6287. doi: 10.3390/s23146287

Algorithm 3: Adaptive Design Algorithm.

[k] = Adaptive(DNN classification results, RF classification results)
Input: CIFAR-100(test set), DNN, RF
Output: optimal parameter k for the secondary RF model
for $k \in {1, 100}$ do
Set all the counters $N_{a}, N_{b}, \dots, N_{f}$ to 0
for image x∈ CIFAR-100 do
if $x \in S E T c r c & D N N (x) \in R F (x, k)$ then
$N_{a} + +$
else
$N_{d} + +$
end if
if $x \in S E T m i s & D N N (x) \notin R F (x, k)$ then
$N_{b} + +$
else
$N_{e} + +$
end if
if $x \in S E T a d v & D N N (x) \notin R F (x, k)$ then
$N_{c} + +$
else
$N_{f} + +$
end if
end for
Objective function $f (k) = (C_{d} \cdot N_{d} + C_{e} \cdot N_{e} + C_{f} \cdot N_{f} - C_{a} \cdot N_{a} - C_{b} \cdot N_{b} - C_{c} \cdot N_{c})$
end for
Among all $f (k), k \in {1, 100}$ find the minimum $f (k^{*})$
Return the optimal index $k^{*}$