Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Achieving Adversarial Robustness by Enforcing Feature Consistency Across Bit Planes
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 |
|---|---|---|---|
| Enforce pre-softmax feature consistency between each image and its stochastically quantized counterpart | ResNet-18 · CIFAR-10, PGD 1000-step white-box | accuracy (%): 0.0 -> 34.4 | Helped |
| Enforce pre-softmax feature consistency between each image and its stochastically quantized counterpart | M-LeNet · Fashion-MNIST, PGD 1000-step white-box | accuracy (%): 0.3 -> 67.7 | Helped |
| Enforce pre-softmax feature consistency between each image and its stochastically quantized counterpart | M-LeNet · MNIST, PGD 1000-step white-box | accuracy (%): 0.0 -> 85.7 | Helped |
| Use BPFC instead of PGD adversarial training | ResNet-50 · ImageNet, targeted PGD 20-step | accuracy (%): 43.43 -> 32.91 | Hurt |
| Remove pre-quantization noise (ablation) | BPFC · CIFAR-10, PGD 1000-step white-box | accuracy (%): 34.4 -> 30.1 | Hurt |
| Replace quantization with uniform random noise (ablation) | BPFC · CIFAR-10, PGD 1000-step white-box | accuracy (%): 34.4 -> 31.9 | Hurt |
| Replace the L2 feature-consistency loss with L1 (ablation) | BPFC · CIFAR-10, PGD 1000-step white-box | accuracy (%): 34.4 -> 35.9; 37.5 -> 24.8 | Mixed |
| Increase the rate of the lambda step-up (ablation) | BPFC on CIFAR-10 · CIFAR-10, clean and adversarial samples | accuracy (%): 75.28; 40.6 | Mixed |
| Add the BPFC loss to the cross-entropy loss during adaptive attack generation (ablation) | BPFC · CIFAR-10, 50-step adaptive PGD | accuracy (%): 34.68 -> 34.52 | No clear effect |
| Add a penalty on high-magnitude LSBs during adaptive attack generation (ablation) | BPFC · CIFAR-10, 50-step adaptive PGD | accuracy (%): 34.68 -> 34.52 | No clear effect |
| Use BPFC rather than normal training | ResNet-18 · CIFAR-10, training time per epoch | seconds: 39.6 -> 69.4 | Hurt |
| Use BPFC rather than PGD adversarial training | ResNet-18 · CIFAR-10, training time per epoch | seconds: 257.8 -> 69.4 | Helped |
Closest papers by meaning
- A Combination of Noise and Bilateral Filters Achieve Supralinear and Scalable Adversarial Robustness in CNNs (CVPR 2026 · similarity 0.90)
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.90)
- Attribute-Guided Adversarial Training for Robustness to Natural Perturbations (AAAI 2021 · similarity 0.90)
- Towards the first adversarially robust neural network model on MNIST (ICLR 2019 · similarity 0.89)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 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.