TransferAttack is a pytorch framework to boost the adversarial transferability for image classification.
Devling into Adversarial Transferability on Image Classification: A Review, Benchmark and Evaluation will be released soon.
We also release a list of papers about transfer-based attacks here.
There are a lot of reasons for TransferAttack, such as:
- A benchmark for evaluating new transfer-based attacks: TransferAttack categorizes existing transfer-based attacks into several types and fairly evaluates various transfer-based attacks under the same setting.
- Evaluate the robustness of deep models: TransferAttack provides a plug-and-play interface to verify the robustness of models, such as CNNs and ViTs.
- A summary of transfer-based attacks: TransferAttack reviews numerous transfer-based attacks, making it easy to get the whole picture of transfer-based attacks for practitioners.
- Python >= 3.6
- PyTorch >= 1.12.1
- Torchvision >= 0.13.1
- timm >= 0.6.12
pip install -r requirements.txtWe randomly sample 1,000 images from ImageNet validate set, in which each image is from one category and can be correctly classified by the adopted models (For some categories, we cannot choose one image that is correctly classified by all the models. In this case, we select the image that receives accurate classifications from the majority of models.). Download the data from 
/path/to/data. Then you can execute the attack as follows:
python main.py --input_dir ./path/to/data --output_dir adv_data/mifgsm/resnet18 --attack mifgsm --model=resnet18
python main.py --input_dir ./path/to/data --output_dir adv_data/mifgsm/resnet18 --eval
| Category | Attack | Main Idea |
|---|---|---|
| Gradient-based | FGSM (Goodfellow et al., 2015) | Add a small perturbation in the direction of gradient |
| I-FGSM (Kurakin et al., 2015) | Iterative version of FGSM | |
| MI-FGSM (Dong et al., 2018) | Integrate the momentum term into the I-FGSM | |
| NI-FGSM (Lin et al., 2020) | Integrate the Nesterov's accelerated gradient into I-FGSM | |
| PI-FGSM (Gao et al., 2020) | Reuse the cut noise and apply a heuristic project strategy to generate patch-wise noise | |
| VMI-FGSM (Wang et al., 2021) | Variance tuning MI-FGSM | |
| VNI-FGSM (Wang et al., 2021) | Variance tuning NI-FGSM | |
| EMI-FGSM (Wang et al., 2021) | Accumulate the gradients of several data points linearly sampled in the direction of previous gradient | |
| AI-FGTM (Zou et al., 2022) | Adopt Adam to adjust the step size and momentum using the tanh function | |
| I-FGS²M (Zhang et al., 2021) | Assigning staircase weights to each interval of the gradient | |
| SMI-FGRM (Han et al., 2023) | Substitute the sign function with data rescaling and use the depth first sampling technique to stabilize the update direction. | |
| VA-I-FGSM (Zhang et al., 2022) | Adopt a larger step size and auxiliary gradients from other categories | |
| RAP (Qin et al., 2022) | Inject the worst-case perturbation when calculating the gradient. | |
| PC-I-FGSM (Wan et al., 2023) | Gradient Prediction-Correction on MI-FGSM | |
| IE-FGSM (Peng et al., 2023) | Integrate anticipatory data point to stabilize the update direction. | |
| GRA (Zhu et al., 2023) | Correct the gradient using the average gradient of several data points sampled in the neighborhood and adjust the update gradient with a decay indicator | |
| GNP (Wu et al., 2023) | Introduce a gradient norm penalty (GNP) term into the loss function | |
| MIG (Ma et al., 2023) | Utilize integrated gradient to steer the generation of adversarial perturbations | |
| DTA (Yang et al., 2023) | Calculate the gradient on several examples using small stepsize | |
| PGN (Ge et al., 2023) | Penalizing gradient norm on the original loss function | |
| NCS (Qiu et al., 2024) | Construct a max-min bi-level optimization problem aimed at finding flat adversarial regions | |
| ANDA (Fang et al., 2024) | Explicitly characterize adversarial perturbations from a learned distribution by taking advantage of the asymptotic normality property of stochastic gradient ascent. | |
| GI-FGSM (Wang et al., 2024) | Use global momentum initialization to better stablize update direction. | |
| Input transformation-based | DIM (Xie et al., 2019) | Random resize and add padding to the input sample |
| TIM (Dong et al., 2019) | Adopt a Gaussian kernel to smooth the gradient before updating the perturbation | |
| SIM (Ling et al., 2020) | Calculate the average gradient of several scaled images | |
| DEM (Zou et al., 2020) | Calculate the average gradient of several DIM's transformed images | |
| Admix (Wang et al., 2021) | Mix up the images from other categories | |
| ATTA (Wu et al., 2021) | Train an adversarial transformation network to perform the input-transformation | |
| MaskBlock (Fan et al., 2022) | Calculate the average gradients of multiple randomly block-level masked images. | |
| SSM (Long et al., 2022) | Randomly scale images and add noise in the frequency domain | |
| AITL (Yuan et al., 2022) | Select the most effective combination of image transformations specific to the input image. | |
| PAM (Zhang et al., 2023) | Mix adversarial examples with base images, where ratios are genreated by a trianed semantic predictor, for gradient accumulation. | |
| LPM (Wei et al., 2023) | Boosting Adversarial Transferability with Learnable Patch-wise Masks | |
| SIA (Wang et al., 2023) | Split the image into blocks and apply various transformations to each block | |
| STM (Ge et al., 2023) | Transform the image using a style transfer network | |
| USMM (Wang et al., 2023) | Apply uniform scale and a mix mask from an image of a different category to the input image | |
| DeCowA (Lin et al., 2024) | Augments input examples via an elastic deformation, to obtain rich local details of the augmented inputs | |
| L2T (Zhu et al., 2024) | Optimizing the input-transformation trajectory along the adversarial iteration | |
| BSR (Wang et al., 2024) | Randomly shuffles and rotates the image blocks | |
| Advanced objective | TAP (Zhou et al., 2018) | Maximize the difference of feature maps between benign sample and adversarial example and smooth the perturbation |
| ILA (Huang et al., 2019) | Enlarge the similarity of feature difference between the original adversarial example and benign sample | |
| ATA (Wu et al., 2020) | Add a regularizer on the difference between attention maps of benign sample and adversarial example | |
| YAILA (Li et al., 2020) | Establishe a linear map between intermediate-level discrepancies and classification loss | |
| FIA (Wang et al., 2021) | Minimize a weighted feature map in the intermediate layer | |
| IR (Wang et al., 2021) | Introduces the interaction regularizer into the objective function to minimize the interaction for better transferability | |
| TRAP (Wang et al., 2021) | Utilize affine transformations and reference feature map | |
| TAIG (Huang et al., 2022) | Adopt the integrated gradient to update perturbation | |
| FMAA (He et al., 2022) | Utilize momentum to calculate the weight matrix in FIA | |
| NAA (Zhang et al., 2022) | Compute the feature importance of each neuron with decomposition on integral | |
| RPA (Zhang et al., 2022) | Calculate the weight matrix in FIA on randomly patch-wise masked images | |
| Fuzziness_Tuned (Yang et al., 2023) | The logits vector is fuzzified using the confidence scaling mechanism and temperature scaling mechanism | |
| DANAA (Jin et al., 2023) | Utilize an adversarial non-linear path to compute feature importance for each neuron by decomposing the integral | |
| ILPD (Li et al., 2023) | Decays the intermediate-level perturbation from the benign features by mixing the features of benign samples and adversarial examples | |
| BFA (Wang et al., 2024) | Calcuate the weight matrix in FIA on adversarial examples generated by I-FGSM | |
| Model-related | SGM (Wu et al., 2020) | Utilize more gradients from the skip connections in the residual blocks |
| LinBP (Guo et al., 2020) | Calculates forward as normal but backpropagates the loss as if no ReLU is encountered in the forward pass | |
| PNA-PatchOut (Wei et al., 2022) | Ignore gradient of attention and randomly drop patches among the perturbation | |
| IAA (Zhu et al., 2022) | Replace ReLU with Softplus and decrease the weight of residual module | |
| SAPR (Zhou et al., 2022) | Randomly permute input tokens at each attention layer | |
| SETR (Naseer et al., 2022) | Ensemble and refine classifiers after each transformer block | |
| DRA (Zhu et al., 2022) | Use fine-tuned models to push the image away from the original distribution while generating the adversarial examples. | |
| MTA (Qin et al., 2023) | Train a meta-surrogate model (MSM), whose adversarial examples can maximize the loss on a single or a set of pre-trained surrogate models | |
| MUP (Yang et al., 2023) | Mask unimportant parameters of surrogate models | |
| TGR (Zhang et al., 2023) | Scale the gradient and mask the maximum or minimum gradient magnitude | |
| DSM (Yang et al., 2022) | Train surrogate models in a knowledge distillation manner and adopt CutMix on the input | |
| DHF (Wang et al., 2023) | Mixup the feature of current examples and benign samples and randomly replaces the features with their means. | |
| BPA (Wang et al., 2023) | Recover the trunctaed gradient of non-linear layers | |
| AGS (Wang et al., 2024) | Train surrogate models with adversary-centric contrastive learning and adversarial invariant learning | |
| MetaSSA (Weng et al., 2024) | Utilizes low-frequency feature mixing for meta-train to compute gradients, averages gradients through adversarial feature mixing during meta-test, and updates adversarial examples using gradients from both steps. | |
| VDC (Zhang et al., 2024) | Adding virtual dense connections for dense gradient back-propagation in Attention maps and MLP blocks, without altering the forward pass. | |
| MA (Ma et al., 2024) | Minimize KL divergence in the predictions between the source and the witness model. | |
| Ensemble-based | Ens (Liu et al., 2017) | Generate the adversarial examplesusing multiple models |
| Ghost (Li et al., 2020) | Densely apply dropout and random scaling on the skip connection to generate several ghost networks to average the gradient | |
| SVRE (Xiong et al., 2020) | Use the stochastic variance reduced gradient to update the adversarial example | |
| LGV (Gubri et al., 2022) | Ensemble multiple weight sets from a few additional training epochs with a constant and high learning rate | |
| MBA (Li et al., 2023) | Maximize the average prediction loss on several models obtained by single run of fine-tuning the surrogate model using Bayes optimization | |
| AdaEA (Chen et al., 2023) | Adjust the weights of each surrogate model in ensemble attack using adjustment strategy and reducing conflicts between surrogate models by reducing disparity of gradients of them | |
| CWA (Chen et al., 2023) | Define the common weakness of an ensemble of models as the solution that is at the flat landscape and close to the models' local optima | |
| SMER (Tang., 2024) | Ensembles reweighing is introduced to refine ensemble weights by maximizing attack loss based on reinforcement learning | |
| Generation-based | CDTP (Naseer et al., 2019) | Train a generative model on datasets from different domains to learn domain-invariant perturbations |
| LTP (Nakka et al., 2021) | Introduce a loss function based on such mid-level features to learn an effective, transferable perturbation generator | |
| ADA (Kim et al., 2022) | Utilize a generator to stochastically perturb shared salient features across models to avoid poor local optima and explore the search space thoroughly | |
| GE-ADVGAN (Zhu et al., 2024) | Enhance the transferability of adversarial samples by incorporating gradient editing mechanisms and frequency domain exploration into the generative model's training process. |
| Category | Attack | Main Idea |
|---|---|---|
| Input transformation-based | ODI (Byun et al., 2022) | Diverse inputs based on 3D objects |
| SU (Wei et al., 2023) | Optimize adversarial perturbation on the original and cropped images by minimizing prediction error and maximizing their feature similarity | |
| IDAA (Liu et al., 2024) | Design local mixup to randomly mix a group of transformed adversarial images, strengthening the input diversity | |
| Advanced objective | AA (Inkawhich et al., 2019) | Minimize the similarity of feature difference between the original adversarial example and target benign sample |
| PoTrip (Li et al., 2020) | Introduce the Poincare distance as the similarity metric to make the magnitude of gradient self-adaptive | |
| Logit (Zhao et al., 2021) | Replace the cross-entropy loss with logit loss | |
| Logit-Margin (Weng et al., 2023) | Downscale the logits using a temperature factor and an adaptive margin | |
| CFM (Byun et al., 2023) | Mix feature maps of adversarial examples with clean feature maps of benign images stocastically | |
| FFT (Zeng et al., 2024) | Fine-tuning a crafted adversarial example in the feature space | |
| Generation-based | TTP (Naseer et al., 2021) | Train a generative model to generate adversarial examples, of which both the global distribution and local neighborhood structure in the latent feature space are matched with the target class. |
| M3D (Zhao et al., 2023) | ||
| Ensemble-based | SASD_WS (Wu et al., 2024) | Incorporate Sharpness-Aware Self-Distillation (SASD) and Weight Scaling (WS) to promote the source model's generalization capability. |
To thoroughly evaluate existing attacks, we have included various popular models, including both CNNs (ResNet-18, ResNet-101, ResNeXt-50, DenseNet-121) and ViTs (ViT, PiT, Visformer, Swin). Moreover, we also adopted four defense methods, namely AT, HGD, RS, NRP. The defense models can be downloaded from Google Drive or Huggingface.
Note: We adopt
| Category | Attacks | CNNs | ViTs | Defenses | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ResNet-18 | ResNet-101 | ResNeXt-50 | DenseNet-101 | ViT | PiT | Visformer | Swin | AT | HGD | RS | NRP | ||
| Gradient-based | FGSM | 96.1 | 33.5 | 36.8 | 60.2 | 15.0 | 17.8 | 26.4 | 32.7 | 33.4 | 25.9 | 22.9 | 29.7 |
| I-FGSM | 100.0 | 14.9 | 18.6 | 42.9 | 4.9 | 10.0 | 14.6 | 21.7 | 30.3 | 8.8 | 20.0 | 13.7 | |
| MI-FGSM | 100.0 | 42.9 | 46.3 | 73.9 | 17.2 | 23.8 | 33.7 | 42.5 | 33.1 | 32.0 | 22.4 | 26.5 | |
| NI-FGSM | 100.0 | 43.8 | 47.2 | 77.0 | 16.6 | 21.5 | 33.3 | 43.2 | 33.0 | 33.2 | 22.5 | 27.3 | |
| PI-FGSM | 100.0 | 37.9 | 46.3 | 72.7 | 14.4 | 17.7 | 27.2 | 37.9 | 37.2 | 37.6 | 31.9 | 36.1 | |
| VMI-FGSM | 100.0 | 62.0 | 64.9 | 88.9 | 28.2 | 39.4 | 53.2 | 58.6 | 36.0 | 53.8 | 26.1 | 40.8 | |
| VNI-FGSM | 100.0 | 62.2 | 64.8 | 89.8 | 26.3 | 35.9 | 52.5 | 56.3 | 34.6 | 50.2 | 25.0 | 38.2 | |
| EMI-FGSM | 100.0 | 57.0 | 59.0 | 89.0 | 21.2 | 28.9 | 44.6 | 52.2 | 35.0 | 43.2 | 24.9 | 32.6 | |
| AI-FGTM | 100.0 | 36.2 | 39.6 | 69.5 | 13.9 | 20.1 | 29.7 | 37.3 | 32.0 | 26.9 | 21.7 | 23.5 | |
| I-FGS²M | 100.0 | 22.7 | 27.0 | 54.5 | 9.0 | 12.1 | 20.1 | 28.9 | 30.8 | 16.2 | 20.2 | 16.6 | |
| SMI-FGRM | 99.8 | 40.2 | 44.5 | 77.1 | 14.0 | 21.0 | 30.7 | 43.9 | 36.6 | 31.6 | 26.0 | 30.5 | |
| VA-I-FGSM | 100.0 | 17.7 | 22.4 | 46.9 | 7.2 | 11.2 | 15.0 | 22.7 | 30.3 | 12.7 | 20.1 | 19.2 | |
| RAP | 100.0 | 51.8 | 58.5 | 87.5 | 21.1 | 26.9 | 43.1 | 49.3 | 32.4 | 39.7 | 22.8 | 31.0 | |
| PC-I-FGSM | 100.0 | 42.8 | 46.8 | 74.5 | 17.1 | 23.6 | 33.4 | 42.8 | 32.9 | 32.1 | 22.9 | 29.3 | |
| IE-FGSM | 100.0 | 51.1 | 54.5 | 83.9 | 19.0 | 28.4 | 40.1 | 47.2 | 33.2 | 39.9 | 22.8 | 28.9 | |
| GRA | 100.0 | 67.9 | 70.0 | 93.9 | 30.3 | 39.3 | 54.5 | 64.2 | 40.8 | 61.0 | 35.1 | 54.8 | |
| GNP | 100.0 | 50.3 | 55.4 | 82.7 | 21.5 | 26.9 | 39.5 | 47.0 | 33.3 | 40.4 | 24.1 | 30.6 | |
| MIG | 100.0 | 54.3 | 58.0 | 87.2 | 22.9 | 31.3 | 44.3 | 53.5 | 37.5 | 47.7 | 26.5 | 39.8 | |
| DTA | 100.0 | 50.6 | 54.8 | 82.5 | 18.1 | 26.0 | 40.2 | 44.8 | 33.0 | 40.6 | 23.1 | 29.2 | |
| PGN | 100.0 | 69.3 | 73.3 | 94.7 | 32.7 | 42.9 | 56.0 | 66.5 | 40.5 | 63.3 | 34.9 | 56.9 | |
| NCS | 100.0 | 80.5 | 82.9 | 96.8 | 45.6 | 54.3 | 68.6 | 75.1 | 41.3 | 74.6 | 36.3 | 62.6 | |
| ANDA | 100.0 | 74.4 | 78.9 | 96.9 | 42.0 | 50.4 | 65.8 | 69.0 | 38.0 | 71.8 | 26.9 | 42.9 | |
| GI-FGSM | 100.0 | 48.0 | 53.6 | 81.7 | 17.8 | 24.9 | 38.3 | 45.4 | 34.0 | 36.9 | 23.7 | 31.2 | |
| Input transformation-based | DIM | 100.0 | 62.7 | 67.3 | 90.3 | 29.5 | 37.1 | 53.7 | 58.7 | 36.5 | 58.5 | 24.9 | 36.0 |
| TIM | 100.0 | 37.2 | 45.0 | 71.8 | 15.5 | 19.6 | 29.3 | 39.1 | 37.4 | 35.2 | 32.5 | 37.4 | |
| SIM | 100.0 | 59.9 | 63.1 | 89.9 | 24.8 | 34.1 | 51.0 | 53.9 | 36.1 | 52.0 | 25.1 | 38.2 | |
| DEM | 100.0 | 76.4 | 78.8 | 97.3 | 39.9 | 45.6 | 66.0 | 67.0 | 38.6 | 78.6 | 30.5 | 47.3 | |
| Admix | 100.0 | 68.2 | 71.8 | 95.1 | 30.0 | 38.6 | 56.1 | 60.5 | 37.6 | 60.1 | 27.6 | 44.2 | |
| ATTA | 100.0 | 46.6 | 50.3 | 79.4 | 17.5 | 26.3 | 37.3 | 45.3 | 33.8 | 38.1 | 22.8 | 30.4 | |
| MaskBlock | 100.0 | 49.2 | 51.4 | 78.6 | 18.0 | 25.1 | 38.1 | 45.6 | 33.9 | 36.8 | 22.9 | 30.5 | |
| SSM | 99.9 | 70.5 | 73.8 | 93.5 | 30.4 | 39.4 | 54.5 | 63.3 | 37.2 | 62.1 | 29.2 | 50.9 | |
| AITL | 99.5 | 78.9 | 82.4 | 96.3 | 46.4 | 51.4 | 68.1 | 71.1 | 41.8 | 79.7 | 32.9 | 53.1 | |
| PAM | 100.0 | 56.5 | 58.5 | 89.1 | 19.7 | 29.7 | 42.8 | 49.9 | 36.3 | 48.0 | 25.0 | 36.0 | |
| LPM | 100.0 | 52.7 | 55.4 | 82.6 | 22.0 | 29.0 | 42.4 | 46.6 | 33.6 | 41.1 | 23.7 | 31.5 | |
| SIA | 100.0 | 87.5 | 90.5 | 99.1 | 43.5 | 57.8 | 77.5 | 78.0 | 39.2 | 81.4 | 28.8 | 51.9 | |
| STM | 100.0 | 75.4 | 77.2 | 96.1 | 35.7 | 45.2 | 61.5 | 68.1 | 40.9 | 70.7 | 32.5 | 58.8 | |
| USMM | 100.0 | 74.0 | 78.1 | 96.4 | 33.7 | 45.3 | 62.8 | 64.8 | 40.0 | 66.1 | 29.4 | 50.8 | |
| DeCowA | 100.0 | 84.8 | 87.7 | 98.6 | 53.6 | 64.0 | 79.5 | 79.7 | 43.6 | 85.7 | 35.2 | 56.0 | |
| L2T | 100.0 | 88.4 | 89.9 | 98.8 | 50.7 | 64.2 | 79.6 | 79.7 | 43.0 | 86.7 | 32.9 | 60.6 | |
| BSR | 100.0 | 85.4 | 87.9 | 99.1 | 42.9 | 56.9 | 74.6 | 77.0 | 38.6 | 80.1 | 27.3 | 48.1 | |
| Advanced objective | TAP | 100.0 | 38.5 | 42.4 | 72.0 | 14.3 | 17.9 | 28.5 | 34.2 | 31.6 | 28.9 | 20.8 | 25.9 |
| ILA | 100.0 | 45.6 | 51.9 | 77.8 | 15.2 | 21.6 | 35.3 | 44.4 | 32.0 | 31.5 | 20.1 | 22.9 | |
| ATA | 100.0 | 16.4 | 19.6 | 41.8 | 5.9 | 8.9 | 14.4 | 21.4 | 30.4 | 10.0 | 20.5 | 15.7 | |
| YAILA | 51.5 | 26.2 | 28.5 | 49.0 | 6.7 | 11.4 | 16.5 | 25.7 | 29.3 | 13.4 | 18.8 | 14.7 | |
| FIA | 99.5 | 31.0 | 36.4 | 65.3 | 10.2 | 16.3 | 24.4 | 35.3 | 31.4 | 18.9 | 21.1 | 19.9 | |
| IR | 100.0 | 42.0 | 45.3 | 74.0 | 16.7 | 23.4 | 33.4 | 40.9 | 40.8 | 32.2 | 28.0 | 22.8 | |
| TRAP | 96.9 | 63.2 | 66.7 | 85.1 | 23.6 | 33.3 | 52.8 | 56.5 | 33.0 | 56.8 | 20.6 | 26.2 | |
| TAIG | 100.0 | 26.0 | 29.1 | 62.0 | 8.4 | 14.1 | 21.8 | 32.4 | 32.3 | 18.3 | 20.9 | 18.2 | |
| FMAA | 100.0 | 39.5 | 44.6 | 80.3 | 11.1 | 20.1 | 29.4 | 41.2 | 32.4 | 25.9 | 21.3 | 22.3 | |
| NAA | 99.5 | 56.5 | 58.9 | 80.8 | 23.9 | 33.9 | 46.8 | 54.5 | 34.8 | 44.2 | 23.9 | 36.8 | |
| RPA | 100.0 | 62.5 | 68.7 | 91.6 | 23.7 | 34.2 | 49.6 | 57.0 | 35.8 | 56.3 | 26.7 | 39.1 | |
| Fuzziness_Tuned | 100.0 | 39.9 | 46.5 | 75.3 | 15.6 | 21.2 | 31.5 | 38.9 | 33.1 | 29.9 | 27.6 | 22.8 | |
| DANAA | 100.0 | 59.6 | 63.8 | 90.4 | 17.3 | 26.4 | 44.7 | 49.8 | 34.8 | 44.9 | 23.4 | 32.4 | |
| ILPD | 70.6 | 68.0 | 68.0 | 72.0 | 31.8 | 46.1 | 52.6 | 55.9 | 33.8 | 50.7 | 24.0 | 50.0 | |
| BFA | 100.0 | 77.8 | 79.7 | 96.7 | 28.9 | 43.4 | 64.2 | 65.9 | 35.9 | 65.6 | 25.7 | 41.0 | |
| Model-related | SGM | 100.0 | 48.4 | 50.9 | 78.5 | 20.1 | 28.7 | 39.7 | 48.3 | 34.9 | 37.5 | 24.2 | 30.9 |
| LinBP | 85.8 | 37.8 | 64.2 | 69.8 | 12.4 | 14.7 | 25.5 | 30.4 | 30.4 | 24.9 | 20.4 | 23.6 | |
| PNA-PatchOut | 47.5 | 34.3 | 36.5 | 45.8 | 81.3 | 39.1 | 40.9 | 53.0 | 31.7 | 29.0 | 22.5 | 27.1 | |
| IAA | 100.0 | 44.2 | 50.6 | 85.1 | 12.8 | 19.6 | 32.8 | 40.4 | 33.3 | 29.4 | 22.0 | 26.0 | |
| SAPR | 66.4 | 50.3 | 53.2 | 65.6 | 96.7 | 57.5 | 60.4 | 75.4 | 35.4 | 41.8 | 24.8 | 31.9 | |
| SETR | 72.6 | 36.6 | 43.4 | 64.5 | 54.3 | 33.6 | 43.5 | 68.8 | 36.5 | 31.6 | 25.5 | 50.7 | |
| DRA | 99.4 | 91.0 | 90.7 | 98.5 | 75.7 | 79.5 | 87.0 | 89.2 | 69.2 | 92.6 | 68.3 | 82.4 | |
| MTA | 82.4 | 44.2 | 46.8 | 74.9 | 12.6 | 17.9 | 31.7 | 41.0 | 30.4 | 34.5 | 19.1 | 19.2 | |
| MUP | 100.0 | 50.7 | 51.0 | 81.2 | 18.5 | 26.3 | 37.4 | 43.3 | 33.8 | 37.1 | 22.7 | 29.6 | |
| TGR | 70.8 | 48.1 | 52.6 | 68.2 | 98.3 | 56.0 | 61.8 | 73.4 | 36.6 | 43.5 | 28.0 | 36.9 | |
| DSM | 98.9 | 60.4 | 66.3 | 91.9 | 23.8 | 33.8 | 49.3 | 56.2 | 34.7 | 48.7 | 24.3 | 34.1 | |
| DHF | 100.0 | 70.4 | 72.1 | 92.3 | 31.5 | 43.4 | 59.8 | 61.9 | 35.9 | 59.8 | 25.5 | 40.2 | |
| BPA | 100.0 | 59.0 | 63.9 | 88.1 | 22.8 | 32.0 | 47.6 | 54.4 | 35.1 | 47.3 | 26.1 | 37.7 | |
| AGS | 86.1 | 55.8 | 60.3 | 81.6 | 29.0 | 22.0 | 46.7 | 46.1 | 37.8 | 62.2 | 27.4 | 39.4 | |
| MetaSSA | 100.0 | 72.8 | 78.3 | 96.0 | 41.3 | 51.3 | 64.3 | 64.8 | 40.7 | 75.4 | 31.5 | 55.1 | |
| VDC | 68.1 | 47.9 | 53.0 | 65.5 | 99.8 | 56.7 | 62.0 | 73.1 | 35.8 | 43.9 | 27.0 | 35.1 | |
| MA | 96.4 | 82.1 | 84.3 | 96.4 | 35.3 | 47.9 | 65.1 | 65.6 | 35.3 | 70.2 | 24.9 | 47.5 | |
| Ensemble-based | ENS | 100.0 | 91.7 | 92.5 | 100.0 | 38.7 | 53.0 | 66.6 | 66.4 | 33.5 | 67.8 | 24.7 | 56.1 |
| Ghost | 64.4 | 93.9 | 63.1 | 66.9 | 19.1 | 29.7 | 39.5 | 42.3 | 31.2 | 36.1 | 21.2 | 54.7 | |
| SVRE | 100.0 | 97.7 | 98.0 | 100.0 | 40.6 | 54.4 | 69.9 | 69.5 | 33.8 | 74.9 | 24.1 | 59.7 | |
| LGV | 97.7 | 69.5 | 69.4 | 93.6 | 23.1 | 29.2 | 43.7 | 51.5 | 34.5 | 52.9 | 24.5 | 37.3 | |
| MBA | 100.0 | 96.0 | 95.2 | 99.8 | 41.9 | 51.8 | 75.1 | 76.8 | 39.5 | 86.1 | 28.7 | 52.1 | |
| AdaEA | 100.0 | 91.9 | 92.7 | 100.0 | 39.4 | 52.4 | 67.3 | 67.0 | 33.9 | 69.6 | 24.3 | 58.0 | |
| CWA | 99.7 | 98.3 | 99.1 | 100.0 | 33.9 | 47.7 | 66.4 | 65.0 | 35.8 | 69.4 | 24.9 | 68.9 | |
| SMER | 89.6 | 77.0 | 86.6 | 89.9 | 41.4 | 57.3 | 71.7 | 69.7 | 38.9 | 70.1 | 26.8 | 69.4 | |
| Generation-based | CDTP | 72.8 | 29.9 | 39.8 | 64.6 | 10.5 | 18.7 | 37.4 | 35.7 | 32.6 | 34.8 | 20.7 | 48.7 |
| LTP | 99.1 | 98.7 | 98.7 | 99.5 | 45.1 | 69.4 | 92.1 | 90.2 | 31.7 | 96.5 | 21.6 | 29.7 | |
| ADA | 69.9 | 47.5 | 45.2 | 63.6 | 8.5 | 11.2 | 31.7 | 29.4 | 29.4 | 37.0 | 20.6 | 16.3 | |
| GE-ADVGAN | 97.7 | 47.7 | 59.8 | 73.4 | 13.8 | 8.2 | 22.6 | 23.2 | 31.1 | 52.5 | 34.0 | 33.0 | |
Note: We adopt labels.csv.
For generation-based targeted attack, TTP and M3D, there are 10 target classes and the class to label mapping is shown below.
Class Number: Class Name
24: Great Grey Owl
99: Goose
245: French Bulldog
344: Hippopotamus
471: Cannon
555: Fire Engine
661: Model T
701: Parachute
802: Snowmobile
919: Street Sign
| Category | Attacks | CNNs | ViTs | Defenses | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ResNet-18 | ResNet-101 | ResNeXt-50 | DenseNet-101 | ViT | PiT | Visformer | Swin | AT | HGD | RS | NRP | ||
| Input transformation-based | ODI | 98.9 | 38.6 | 45.5 | 67.0 | 9.4 | 13.9 | 29.5 | 19.9 | 0.1 | 41.4 | 0.0 | 1.0 |
| SU | 99.2 | 7.2 | 8.0 | 19.7 | 0.1 | 0.6 | 2.1 | 1.8 | 0.1 | 2.1 | 0.0 | 0.2 | |
| IDAA | 87.1 | 2.6 | 3.0 | 13.0 | 1.3 | 1.8 | 2.1 | 3.3 | 0.4 | 1.5 | 0.0 | 0.1 | |
| Advanced objective | AA | 5.0 | 0.7 | 0.7 | 0.9 | 0.3 | 0.3 | 0.3 | 0.1 | 0.0 | 0.2 | 0.0 | 0.0 |
| PoTrip | 100.0 | 3.2 | 5.1 | 15.7 | 0.1 | 0.3 | 1.3 | 1.1 | 0.0 | 3.0 | 0.0 | 0.2 | |
| Logit | 99.0 | 13.5 | 18.5 | 38.5 | 1.9 | 2.9 | 8.3 | 3.8 | 0.1 | 14.4 | 0.0 | 0.3 | |
| Logit-Margin | 100.0 | 13.6 | 19.1 | 42.8 | 1.8 | 3.3 | 8.4 | 4.4 | 0.0 | 14.0 | 0.0 | 0.2 | |
| CFM | 98.3 | 39.6 | 44.8 | 66.1 | 9.6 | 11.4 | 26.6 | 18.9 | 0.2 | 37.6 | 0.0 | 1.6 | |
| FFT | 99.8 | 17.5 | 21.6 | 45.1 | 1.3 | 2.8 | 10.3 | 6.6 | 0.1 | 13.2 | 0.0 | 0.4 | |
| Generation-based | TTP | 96.2 | 19.6 | 27.4 | 62.4 | 3.2 | 4.3 | 19.5 | 5.3 | 0.0 | 0.0 | 0.3 | 4.1 |
| M3D | |||||||||||||
| Ensemble-based | SASD_WS | 91.7 | 70.9 | 76.9 | 91.5 | 13.7 | 22.5 | 39.9 | 29.0 | 0.1 | 64.7 | 0.1 | 5.7 |
Xiaosen Wang |
Zeyuan Yin |
Zeliang Zhang |
Kunyu Wang |
Zhijin Ge |
Yuyang Luo |
We thank all the researchers who contribute or check the methods. See contributors for details.
We are trying to include more transfer-based attacks. We welcome suggestions and contributions! Submit an issue or pull request and we will try our best to respond in a timely manner.