Computer Science > Computer Vision and Pattern Recognition
[Submitted on 9 Mar 2023 (v1), last revised 14 Aug 2023 (this version, v2)]
Title:Decision-BADGE: Decision-based Adversarial Batch Attack with Directional Gradient Estimation
View PDFAbstract:The susceptibility of deep neural networks (DNNs) to adversarial examples has prompted an increase in the deployment of adversarial attacks. Image-agnostic universal adversarial perturbations (UAPs) are much more threatening, but many limitations exist to implementing UAPs in real-world scenarios where only binary decisions are returned. In this research, we propose Decision-BADGE, a novel method to craft universal adversarial perturbations for executing decision-based black-box attacks. To optimize perturbation with decisions, we addressed two challenges, namely the magnitude and the direction of the gradient. First, we use batch loss, differences from distributions of ground truth, and accumulating decisions in batches to determine the magnitude of the gradient. This magnitude is applied in the direction of the revised simultaneous perturbation stochastic approximation (SPSA) to update the perturbation. This simple yet efficient method can be easily extended to score-based attacks as well as targeted attacks. Experimental validation across multiple victim models demonstrates that the Decision-BADGE outperforms existing attack methods, even image-specific and score-based attacks. In particular, our proposed method shows a superior success rate with less training time. The research also shows that Decision-BADGE can successfully deceive unseen victim models and accurately target specific classes.
Submission history
From: Geunhyeok Yu [view email][v1] Thu, 9 Mar 2023 01:42:43 UTC (921 KB)
[v2] Mon, 14 Aug 2023 08:08:50 UTC (14,433 KB)
Ancillary-file links:
Ancillary files (details):
- code/README.md
- code/cfg/__init__.py
- code/cfg/aug/cifar10.yaml
- code/cfg/cmd/run/default.yaml
- code/cfg/cmd/stream/cifar10-.yaml
- code/cfg/cmd/stream/default.yaml
- code/cfg/cmd/stream/mnist-budgets.yaml
- code/cfg/train_attack/cifar10-optim/adam.yaml
- code/cfg/train_attack/cifar10-optim/momentum.yaml
- code/cfg/train_attack/cifar10-optim/nag.yaml
- code/cfg/train_attack/cifar10-optim/sgd.yaml
- code/cfg/train_attack/cifar10-toy.yaml
- code/cfg/train_attack/cifar10-vgg11.yaml
- code/cfg/train_attack/default.yaml
- code/cfg/train_attack/mnist-optim/adam.yaml
- code/cfg/train_attack/mnist.yaml
- code/cfg/train_attack/no-scheduling.yaml
- code/cfg/train_attack_imagenet/default.yaml
- code/data/__init__.py
- code/data/augmentations.py
- code/data/imagenet1k_categories.txt
- code/data/imagenet1k_labels/data.pkl
- code/data/imagenet1k_labels/version
- code/data/samplers.py
- code/evaluation.py
- code/hyp_test.ipynb
- code/loss_func.py
- code/make_decision.py
- code/models/__init__.py
- code/models/densenet.py
- code/models/dla.py
- code/models/dla_simple.py
- code/models/dpn.py
- code/models/efficientnet.py
- code/models/googlenet.py
- code/models/lenet.py
- code/models/mnist.py
- code/models/mobilenet.py
- code/models/mobilenetv2.py
- code/models/pnasnet.py
- code/models/preact_resnet.py
- code/models/regnet.py
- code/models/resnet.py
- code/models/resnet_cifar.py
- code/models/resnext.py
- code/models/senet.py
- code/models/shufflenet.py
- code/models/shufflenetv2.py
- code/models/swin.py
- code/models/toy.py
- code/models/vgg.py
- code/models/vit.py
- code/models/vit_small.py
- code/optimizers/__init__.py
- code/recipe/_temp.sh
- code/recipe/cifar10-batchsize_time.sh
- code/recipe/cifar10-budget.sh
- code/recipe/cifar10-hyp_test.sh
- code/recipe/cifar10-resnet18-optim.sh
- code/recipe/cifar10-resnet18-targeted.sh
- code/recipe/imagenet_tune.sh
- code/recipe/l2_budget.sh
- code/recipe/loss.sh
- code/recipe/spsa_batchsize_budget.sh
- code/recipe/spsa_cifar10_budget.sh
- code/recipe/spsa_hyp_test_algs.sh
- code/recipe/spsa_hyp_test_dsets.sh
- code/recipe/spsa_loss.sh
- code/recipe/spsa_mnist_budget.sh
- code/recipe/spsa_optim.sh
- code/recipe/spsa_targeted.sh
- code/recipe/spsa_transformers_cifar10_5000.sh
- code/recipe/targeted_attack.sh
- code/schedulers/__init__.py
- code/train_attack.py
- code/train_attack_detailed.py
- code/train_attack_nes.py
- code/train_attack_spsa.py
- code/train_victim.py
- code/utils.py
- code/vis_algorithms.ipynb
- code/vis_decision.py
- code/vis_gif_dep.py
- code/vis_gif_uap.py
- code/vis_targeted.py
- code/vis_transferability.py
- code/vis_transferability2.py
- demo.mp4
References & Citations
Bibliographic and Citation Tools
Bibliographic Explorer (What is the Explorer?)
Connected Papers (What is Connected Papers?)
Litmaps (What is Litmaps?)
scite Smart Citations (What are Smart Citations?)
Code, Data and Media Associated with this Article
alphaXiv (What is alphaXiv?)
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub (What is DagsHub?)
Gotit.pub (What is GotitPub?)
Hugging Face (What is Huggingface?)
Papers with Code (What is Papers with Code?)
ScienceCast (What is ScienceCast?)
Demos
Recommenders and Search Tools
Influence Flower (What are Influence Flowers?)
CORE Recommender (What is CORE?)
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.