Research map / Adversarial robustness and security / Black-box and transfer attacks
AutoZOOM: Autoencoder-Based Zeroth Order Optimization Method for Attacking Black-Box Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use random full-gradient estimation and a reduced-dimensional decoder instead of coordinate-wise ZOO gradients | ZOO black-box attack · MNIST, CIFAR-10, and ImageNet targeted attacks | mean query count (initial success): 16,533.30 -> 729.65; 11,662.80 -> 332.43; 2,226,405.04 (2.22M) -> 13,525.00 | Helped |
| Use an autoencoder decoder rather than bilinear resizing | AutoZOOM · CIFAR-10 targeted, λini=1 | mean query count (initial success): 813.01 -> 332.43 | Helped |
| Use an autoencoder decoder rather than bilinear resizing | AutoZOOM · MNIST targeted, λini=1 | mean query count (initial success): 879.98 -> 729.65 | Helped |
| Use an autoencoder decoder rather than bilinear resizing | AutoZOOM · ImageNet targeted | mean query count (initial success): 14,228.88 -> 13,525.00 | Helped |
| Increase post-success averaging parameter q above 1 (ablation) | AutoZOOM-AE · ImageNet targeted post-success refinement | distortion | Helped |
| Increase post-success averaging parameter beyond q=4 (ablation) | AutoZOOM-AE · ImageNet targeted post-success refinement | distortion | No clear effect |
| Reduce the attack dimension using the autoencoder decoder (ablation) | AutoZOOM-AE · MNIST and CIFAR-10 targeted attacks | mean query count: roughly 35-40% | Helped |
| Reduce the attack dimension using the autoencoder decoder (ablation) | AutoZOOM-AE · ImageNet targeted attack | mean query count: at least 95% | Helped |
| Use bilinear resizing rather than the autoencoder decoder (ablation) | AutoZOOM · CIFAR-10 targeted, λini=10 | queries to reach the same distortion: 835.27 vs 990.61 | Mixed |
| Set λini to a larger value (ablation) | AutoZOOM-AE · MNIST and CIFAR-10 targeted attacks | mean query count (initial success): MNIST 2,428.24 -> 729.65 -> 510.38; CIFAR-10 1,523.91 -> 332.43 -> 259.34 | Helped |
| Use random full-gradient estimation rather than coordinate-wise gradient estimation | ZOO · MNIST, CIFAR-10, and ImageNet targeted attacks | mean query count (initial success): 16,533.30 -> 729.65; 11,662.80 -> 332.43; 2,226,405.04 (2.22M) -> 13,525.00 | Helped |
| Use AutoZOOM-BiLIN instead of ZOO | ZOO · MNIST, CIFAR-10, and ImageNet untargeted attacks | mean query count (initial success): 7856.64 -> 98.82; 3957.76 -> 85.6; 271627.91 -> 1695.27 | Helped |
Closest papers by meaning
- Simple Black-box Adversarial Attacks (ICML 2019 · similarity 0.88)
- Towards Query-Efficient Black-Box Adversary with Zeroth-Order Natural Gradient Descent (AAAI 2020 · similarity 0.87)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.86)
- DifAttack: Query-Efficient Black-Box Adversarial Attack via Disentangled Feature Space (AAAI 2024 · similarity 0.85)
- BRUSLEATTACK: A QUERY-EFFICIENT SCORE- BASED BLACK-BOX SPARSE ADVERSARIAL ATTACK (ICLR 2024 · similarity 0.85)
- Efficient Black-box Adversarial Attacks via Bayesian Optimization Guided by a Function Prior (ICML 2024 · similarity 0.85)
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.