Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Examples Detection in Deep Networks With Convolutional Filter Statistics
Xin Li, Fuxin Li
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 |
|---|---|---|---|
| use cascade over layers instead of the best single layer (ablation) | detector (AlexNet) · Normal vs L-BFGS adversarials | detection accuracy (%): 78.95 -> 83.4 | Improved |
| use cascade over layers instead of the best single layer (ablation) | detector (VGG-16) · Normal vs L-BFGS adversarials | detection accuracy (%): 84.1 -> 90.665 | Improved |
| use deeper VGG-16 backbone instead of AlexNet for the cascade (ablation) | cascade detector · Normal vs L-BFGS adversarials | detection accuracy (%): 83.4 -> 90.665 | Improved |
| use cascade statistics detector instead of OpenMax | detector (AlexNet) · Normal vs L-BFGS adversarials | accuracy / AUC: outperform OpenMax by over 9% in AUC and 11% in accuracy | Improved |
| detect EA adversarials with a detector trained only on L-BFGS adversarials | cascade detector (AlexNet) · Normal vs EA adversarials | accuracy / AUC: 97.3% accuracy, 98.2% AUC | Improved |
| use first three conv layers only for EA detection (ablation) | single-layer statistics (AlexNet) · Normal vs EA adversarials | detection accuracy (%): 93.45 ± 0.69 -> 98.3 ± 0.73 | Improved |
| add 3x3 average filter to detected adversarial images before classification | VGG-16 classifier · Recovery of detected adversarials | top-5 accuracy: almost 0% -> 73.0% | Improved |
| use 5x5 instead of 3x3 average filter for recovery (ablation) | recovery filter · Recovery of detected adversarials | top-5 accuracy: 73.0% -> 68.0% | Hurt |
| use foveation crop instead of average filter for recovery | recovery · Recovery of detected adversarials | top-5 accuracy: 73.0% -> 82.6% | Improved |
| use PCA tail-distribution features as detector | tail-statistics classifier · Normal vs L-BFGS adversarials | detection accuracy (%): >99% accuracy | Mixed |
| apply detector to DeepFool adversarials on ResNet-50 | cascade detector · Normal vs DeepFool adversarials | AUC: more than 90% AUC | Improved |
| add first-conv-layer prefiltering decision threshold at 97% (AlexNet) / 98% (VGG) true positive rate | cascade detector · Normal vs adversarials | true positive rate: 97% and 98% | No clear effect |
Closest papers by meaning
- Detecting Adversarial Examples Through Image Transformation (AAAI 2018 · similarity 0.90)
- Data ID Extraction Networks for Unsupervised Class- and Classifier-Free Detection of Adversarial Examples (TPAMI 2025 · similarity 0.89)
- Beating Attackers At Their Own Games: Adversarial Example Detection Using Adversarial Gradient Directions (AAAI 2021 · similarity 0.89)
- Natural Adversarial Examples (CVPR 2021 · similarity 0.89)
- GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification (ICLR 2020 · similarity 0.89)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.86)
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.
