Research map / Adversarial robustness and security / Attacks on text and graphs
Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples
Check 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 |
|---|---|---|---|
| replace one-pass feature distillation with two-pass feature distillation (ablation) | Feature Distillation on MobileNet · ImageNet, gray-box (six attacks average) | accuracy on adversarial examples (%): 56% v.s. 91% | Helped |
| use two-pass feature distillation instead of one-pass feature distillation (ablation) | Feature Distillation · ImageNet, gray-box, FGSM | accuracy on adversarial examples (%): 13 -> 92 | Helped |
| use two-pass feature distillation instead of one-pass feature distillation (ablation) | Feature Distillation · ImageNet, gray-box, BIM | accuracy on adversarial examples (%): 35 -> 99 | Helped |
| use two-pass feature distillation instead of one-pass feature distillation (ablation) | Feature Distillation · ImageNet, gray-box, DeepFool | accuracy on adversarial examples (%): 63 -> 99 | Helped |
| use two-pass feature distillation instead of one-pass feature distillation (ablation) | Feature Distillation · ImageNet, gray-box, CW0 | accuracy on adversarial examples (%): 61 -> 58 | Hurt |
| increase the reference quantization step to 2 times (ablation) | Feature Distillation · ImageNet, BPDA | accuracy on adversarial examples (%): 10 -> 42 | Helped |
| increase the reference quantization step to 3 times (ablation) | Feature Distillation · ImageNet, BPDA | accuracy on adversarial examples (%): 10 -> 60 | Helped |
| increase the reference quantization step to 3 times (ablation) | Feature Distillation · ImageNet, BPDA | benign-image accuracy (acc_l, %): 77 -> 74 | Hurt |
| use FD-2P instead of no defense | MobileNet · ImageNet, gray-box (six attacks average) | accuracy on adversarial examples (%): 3.5 -> 91 | Helped |
| use FD-2P instead of TVM | MobileNet · ImageNet, gray-box (six attacks average) | accuracy on adversarial examples (%): 59.8 -> 91 | Helped |
| use FD (3x) instead of JPEG (10) | Inception v3 · ImageNet, BPDA | accuracy on adversarial examples (%): 45 -> 60 | Helped |
| use FD (3x) instead of JPEG (10) | Inception v3 · ImageNet, BPDA | benign-image accuracy (acc_l, %): 61 -> 74 | Helped |
Closest papers by meaning
- JPEG Inspired Deep Learning (ICLR 2025 · similarity 0.86)
- Conditional Backdoor Attack via JPEG Compression (AAAI 2024 · similarity 0.85)
- Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature (AAAI 2025 · similarity 0.85)
- Robust Overfitting Does Matter: Test-Time Adversarial Purification With FGSM (CVPR 2024 · similarity 0.85)
- ROME is Forged in Adversity: Robust Distilled Datasets via Information Bottleneck (ICML 2025 · similarity 0.85)
- BlurGuard: A Simple Approach for Robustifying Image Protection Against AI-Powered Editing (NeurIPS 2025 · similarity 0.84)
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.