Awesome

Poison-adv-training

This code repository provides two poison attack methods (targeted attack & untargeted attack) against adversarial training algorithms. It aims to assess and reveal the vulnerabilities of adversarial learning algorithms. Paper link: https://arxiv.org/abs/2305.00399

Overview

Adversarial training (AT) is a robust learning algorithm that can defend against adversarial attacks in the inference phase and mitigate the side effects of corrupted data in the training phase. As such, it has become an indispensable component of many artificial intelligence (AI) systems. However, in high-stake AI applications, it is crucial to understand AT's vulnerabilities to ensure reliable deployment.

In this work, we investigate AT's susceptibility to poisoning attacks, a type of malicious attack that manipulates training data to compromise the performance of the trained model. Previous work has focused on poisoning attacks against standard training, but little research has been done on their effectiveness against AT. To fill this gap, we design and test effective poisoning attacks against AT.

Specifically, we investigate and design clean-label poisoning attacks, allowing attackers to imperceptibly modify a small fraction of training data to control the algorithm's behavior on a specific target data point. Additionally, we propose the clean-label untargeted attack, enabling attackers can attach tiny stickers on training data to degrade the algorithm's performance on all test data, where the stickers could serve as a signal against unauthorized data collection.

We separately illustrate the two types of poison attacks mentioned above in Figure 1.

Prerequisites

Python (3.8)
Pytorch (1.8.0)
CUDA
Numpy

Usage

Targeted Attack

First, in Figure 2, we provides a visualization of clean-label targeted poisoning attacks carried out by Witches' Brew (WB) [link] and our proposed targeted attack method, respectively. The top row shows unperturbed natural images, while the second and third rows depict the imperceptible noise (scaled by 3 times for visualization) and the poisoned images by WB. Similarly, the fourth and fifth rows depict the imperceptible noise (scaled by 3 times for visualization) and the poisoned images generated by our method. All poisoning noises are bounded by $\ell_{\infty}$-norm with $\epsilon = 16/255$.

To generate clean-label targeted poison data and validate it on the robust learner, you can simply run the following command.

python tar_poisoning.py --dataset CIFAR10 --poisonkey 2000000000 --eps 16 --tau 0.01 --budget 0.04 --attackiter 250 --restarts 8 --vruns 4

If you want to skip the generation stage and directly validate the poison attack, we have also shared the targeted poison data we generated on Google Drive. You can use the validate_tar_poisoning.py to load and evaluate its effectiveness on the robust learner.

Here, we present our evaluation results for WB and our proposed attack on natural (Nat.) and adversarial (Adv.) targets, as shown in Figure 3. Note that a lower Log-scaled logit difference indicates a higher success rate of the attacks. When it falls below 0 (dashed line), it indicates that the majority of the attacks have been successful. The default dataset is CIFAR-10. You can specify the experimental dataset as Tiny ImageNet by using the parameter --dataset TinyImageNet. To reproduce the results shown in Figure 3, you can adjust the parameter --eps from $4$ to $16$.

Untargeted Attack

Similarly, as shown in Figure 4, we first demonstrate the visualization of clean-label untargeted poisoning attacks generated by robust error-minimizing noise (REM) [link] and our method. The top row displays the original natural images. The second and third rows show the imperceptible noise (scaled by 31 times for visualization) and the corresponding poisoned images generated by REM. The fourth and fifth rows display the human-visible stickers and our poisoned images.

On the ImageNet Subset, we visualize the clean-label untargeted poisoning attacks performed by our sticker in Figure 5. The first row has unperturbed natural images. The second row has our poisoned images attached with stickers.

Run the following command to generate clean-label untargeted poison data. For convenience, we have shared our generated untargeted poison data on Google Drive. You can download it directly to skip the generation step.

python untar_poisoning.py --dataset cifar10 --train-steps 5000 --optim sgd --lr 0.1 --lr-decay-rate 0.1 --lr-decay-freq 2000 --pgd-steps 10 --pgd-step-size 35 --pgd-random-start --patch-size 0.03 --location 4

Then, run the following command to validate clean-label untargeted poison data on the robust learner.

python validate_untar_poisoning.py --seed 2000000000 --noise-rate 0.5 --train-steps 15000 --optim sgd --lr 0.1 --lr-decay-rate 0.1 --lr-decay-freq 6000 --pgd-radius 8 --pgd-steps 10 --pgd-step-size 1.6 --pgd-random-start --report-freq 200 --save-freq 100000 --noise-path ./exp_data/untargeted/patch_cifar10_loc4_ss35_ps0.03/poisons/patch-fin-def-noise.pkl --mask-path ./exp_data/untargeted/patch_cifar10_loc4_ss35_ps0.03/poisons/patch-fin-def-mask.pkl --save-name train

In Table 1, we report the evaluation results of robust learners trained with different perturbation radii $\epsilon_{0}$ on both clean training set and poisoned training sets. The poisoned training sets are generated by REM and our untargeted attack method, respectively. For clarity, we have used '+' or '-' inside parentheses to indicate whether the poison attacks improved or decreased the performance of the robust learner, respectively. A greater decrease indicates a more effective attack.

You can replicate the results in the table by modifying the parameters in the validation script validate_untar_poisoning.py. Use --dataset to modify the experimental dataset, --pgd-radius and --pgd-step-size to adjust the training perturbation radius and step size of the robust learner, and --noise-path and --mask-path to specify the paths for the noise and mask data, respectively.

For the ImageNet Subset, in Figure 6, we present the test accuracy curves of the robust learner trained with a perturbation radius of $\epsilon_{0} = 8/255$ on both the clean training set and the poisoned training set generated by our untargeted attack.