Clean Test Accuracy and Adversarial Training via Average Attack | HackerNoon
Briefly

from Hackernoon1 year ago
NEO-KD maintains similar clean accuracy levels to Adv. w/o Distill on Tiny-ImageNet, while significantly boosting adversarial test accuracy, indicating its effectiveness.
Hackernoonhttps://hackernoon.com/clean-test-accuracy-and-adversarial-training-via-average-attack
The Adversarial Training via Average Attack section emphasizes that NEO-KD demonstrates advantages in adversarial training, particularly validated through experiments on datasets like CIFAR-100.
Hackernoonhttps://hackernoon.com/clean-test-accuracy-and-adversarial-training-via-average-attack
Read at Hackernoon
#adversarial-training #accuracy #neo-kd #tiny-imagenet #cifar-100
[
|
]