Adversarial Defense
The Adversarial Defense task in the TrojAI competition aims to develop and evaluate machine learning models that can resist adversarial attacks. The goal of this task is to enhance the robustness and security of the models when faced with malicious inputs, ensuring their reliability and stability in real-world applications. By optimizing defense strategies, the ability of the models to identify and resist potential threats is strengthened, thereby improving the overall security protection level of the system.
CIFAR-10
Stochastic-LWTA/PGD/WideResNet-34-10
ImageNet (non-targeted PGD, max perturbation=4)
SAT-EfficientNet-L1
CIFAR-100
resnet18
ImageNet
ResNet101
ImageNet (targeted PGD, max perturbation=16)
ResNet-152 Denoise
MNIST
Defense GAN
CAAD 2018
Feature Denoising
Auto Encoder-Block Switching defense with GradCAM
TrojAI Round 0
TrojAI Round 1