Open Access
Review
Table 3
Representative adversarial defense methods
Defense method | Year | Venue | Applicable DNNs | Description |
---|---|---|---|---|
Adversarial training (AT) | ||||
PGD-AT[6] | 2018 | ICLR | ResNet | Using PGD for AT |
TRADES[8] | 2019 | ICML | CNN, ResNet-18 | Trade-off in AT between robustness and accuracy |
MART[63] | 2019 | ICLR | ResNet-18 | Improving AT by revisiting misclassified samples |
FAT[9] | 2020 | ICML | CNN, ResNet | Friendly AT with an early stopping mechanism |
Fast AT[64] | 2020 | ICLR | ResNet-18, ResNet-50 | Using PGD for adversarial training |
Patch-based negative augmentation[76] | 2022 | NeurIPS | ViT | Introducing negatively enhanced samples to train the model |
LAS-AT[11] | 2022 | CVPR | ResNet-18, WideResNet-34-10 and PreActResNet-18 |
Adaptive AT with a strategy network |
UIAT[77] | 2023 | CVPR | ResNet-18, PreActResNet-18 and WideResNet-28-10 | Improving AT with inverse adversarial examples |
FP-Better[78] | 2024 | TIFS | ResNet18 | Fast propagation boosted AT |
Robust network design | ||||
FD[65] | 2019 | CVPR | ResNet-101, ResNet-152 | Feature denoising between network layers |
ANP-VS[66] | 2020 | ICML | LeNet 5-Caffe, VGG-16 | Adversarial neural pruning |
FSR[67] | 2023 | CVPR | ResNet-18, VGG-16 and WideResNet-34-10 | Feature separation and recalibration modules |
PAAS[79] | 2024 | CVPR | ViT | Changing the attention mechanism within the network |
Input transformation | ||||
MagNet[80] | 2017 | SIGSAC | CNN | Preprocessing with denoising auto-encoders |
Quilting[71] | 2018 | ICLR | ResNet-50 | Piecing together clean images using a database |
TVM[71] | 2018 | ICLR | ResNet-50 | Total variance minimization |
Defense-GAN[12] | 2018 | ICLR | CNN | Adversarial Purification (AP): GANs |
ComDefend[81] | 2019 | CVPR | ComCNN + RecCNN | Bit depth compression and reconstruction |
BuZz[49] | 2020 | ArXiv | ResNet-56,VGG-16 and ViT | Apply a unique transformation to the input image |
A-VAE[82] | 2020 | ECCV | ResNet-50, VGGFace2 | AP: combining VAEs and GANs |
MCMC-EBM[72] | 2021 | ICLR | ResNet-50 | AP: Monte-Carlo Markov-Chain energy-based model |
DiffPure[10] | 2022 | ICML | ResNet-50, WideResNet-50-2 and ViT | AP: deep diffusion probabilistic models |
DISCO[83] | 2022 | NeurIPS | ResNet-18 and WideResNet-28 | AP: super-resolution algorithms |
Certified defenses | ||||
Safety verification[68] | 2017 | CAV | CNN | Verification with satisfiability modulo theory |
COAP[84] | 2018 | ICML | CNN | Defense with convex outer adversarial polytope |
IBP[73] | 2018 | NeurIPS | CNN | Defense with interval bound propagation |
Lipschitz continuity[69] | 2019 | NeurIPS | CNN | Use Lipschitz constant to bound the change in the output |
MILP[75] | 2019 | ICLR | CNN | Mixed integer linear programming for robustness evalaution |
PCT[85] | 2020 | ICLR | U-Net | Certified robustness against adversarial patches |
Certifiable patch defense[86] | 2022 | CVPR | ResNet-50, ResNext-10 and ViT-S/16-224 | Practical verifiable patch defense |
Ensemble defenses | ||||
EAT[70] | 2018 | ICLR | InceptionResNet-v2 | AT with an ensemble of models |
Diversity[87] | 2019 | ICML | ResNet-20 | Improving model diversity for ensemble defense |
P&D[88] | 2021 | AAAI | CNN, ResNet | Improving model diversity and reducing transferability |
SEAT[89] | 2022 | ICLR | ResNet-18 and WideResNet-32-10 | Self-ensemble AT using history models |
iGAT[90] | 2023 | NeurIPS | ResNet-20 | A theory of ensemble defenses and improved methods |
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.