Research map / Adversarial robustness and security / Black-box and transfer attacks
Practical Black-box Attacks on Deep Neural Networks using Efficient Query Mechanisms
Arjun Nitin Bhagoji, Warren He, Bo Li, Dawn Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add random grouping (RG-8) query reduction to iterative FD attack (ablation) | IFD-logit, Model A · MNIST targeted (eps=0.3) | targeted attack success rate (%): 99.7 -> 73.8 | Hurt |
| add random grouping (RG-8) query reduction to iterative FD attack (ablation) | IFD-logit, Resnet-32 · CIFAR-10 targeted (eps=8) | targeted attack success rate (%): 100.0 -> 97.0 | Mixed |
| use PCA-100 instead of random grouping for single-step query reduction (ablation) | Gradient Estimation, Model A · MNIST targeted (eps=0.3) single-step | targeted attack success rate (%): 15.9 -> 23.2 | Improved |
| use PCA-100 instead of random grouping for iterative query reduction (ablation) | Gradient Estimation, Model A · MNIST targeted (eps=0.3) | targeted attack success rate (%): 73.8 -> 96.2 | Improved |
| use iterative instead of single-step finite-difference attack | FD-logit, Model A · MNIST targeted (eps=0.3) | targeted attack success rate (%): 29.9 -> 99.7 | Improved |
| add random grouping (RG-8) to single-step finite-difference attack (ablation) | Finite Diff., Model A · MNIST untargeted (eps=0.3) | attack success rate (%): 92.9 -> 61.5 | Hurt |
| use iterative Gradient Estimation RG-8 instead of SPSA | Gradient Estimation, Model A · MNIST untargeted (eps=0.3) | attack success rate (%): 96.7 -> 98.4 | Improved |
| use iterative Gradient Estimation RG-8 instead of SPSA | Gradient Estimation, Resnet-32 · CIFAR-10 untargeted (eps=8) | attack success rate (%): 88.0 -> 99.0 | Improved |
| use Particle Swarm Optimization instead of single-step Finite Diff. | PSO, Model A · MNIST untargeted (eps=0.3) | attack success rate (%): 92.9 -> 84.1 | Hurt |
| train target with iterative adversarial training instead of FGS adversarial training | IFD-logit, Model A · MNIST untargeted (eps=0.3) | untargeted attack success rate (%): 76.5 -> 11.6 | Hurt |
| train target with iterative adversarial training instead of FGS adversarial training | IFD-logit, Resnet-32 · CIFAR-10 adversarially trained (eps=8) | untargeted attack success rate (%): 100.0 -> 54.6 | Hurt |
| round output probabilities to two decimal places as countermeasure (ablation) | Gradient Estimation attacks · MNIST/CIFAR-10 undefended | attack success rate (%): 28.0% | Hurt |
Closest papers by meaning
- Improving Query Efficiency of Black-box Adversarial Attack (ECCV 2020 · similarity 0.92)
- Towards Query-Efficient Black-Box Adversary with Zeroth-Order Natural Gradient Descent (AAAI 2020 · similarity 0.92)
- Query-Efficient Black-Box Adversarial Attacks Guided by a Transfer-Based Prior (TPAMI 2021 · similarity 0.92)
- Query-Efficient Black-Box Adversarial Attack With Customized Iteration and Sampling (TPAMI 2022 · similarity 0.92)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.91)
- Efficient Black-box Adversarial Attacks via Bayesian Optimization Guided by a Function Prior (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
