Research map / Adversarial robustness and security / Black-box and transfer attacks
Attacking Binarized Neural Networks
Angus Galloway, Graham W. Taylor, Medhat Moussa
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 binary rather than full-precision weights and activations without adversarial training | MNIST CNN, 256 first-layer kernels · MNIST test, FGSM ε=0.3 | accuracy: 17±3% -> 78±1% | Improved |
| Add a learned scalar after hidden-layer ReLUs to the binary model (ablation) | MNIST CNN, 256 first-layer kernels · MNIST clean test | error rate: 2.3±0.2% -> 1.2±0.1% | Improved |
| Train the scaled binary model with PGD rather than no adversarial training (ablation) | MNIST CNN, 256 first-layer kernels · MNIST test, FGSM ε=0.3 | accuracy: 16±3% -> 85.6±0.6% | Improved |
| Add PGD adversarial training to a full-precision model (ablation) | MNIST CNN, 256 first-layer kernels · MNIST test, FGSM ε=0.3 | accuracy: 17±3% -> 85±1% | Improved |
| Stochastically quantize weights instead of deterministic binary quantization (ablation) | MNIST binary CNN · MNIST targeted CW L2, 100 iterations | accuracy: 22±3% -> 71±2% | Improved |
| Increase binary-model capacity from 32 to 256 first-layer kernels (ablation) | MNIST binary CNN · MNIST targeted CW L2, 100 iterations | accuracy: 7±1% -> 22±3% | Improved |
| Add FGSM adversarial training to deterministic binary models (ablation) | MNIST binary CNN · MNIST test, FGSM ε=0.3 | accuracy: 78±1% -> 58±2% | Hurt |
| Add FGSM adversarial training to scaled binary models (ablation) | CIFAR-10 CNN, 256 first-layer kernels · CIFAR-10 black-box transfer, FGSM ε=0.3 | transfer accuracy: 21±1% -> 27.6±0.9% | Mixed |
| Use PGD training rather than FGSM training on the scaled binary model (ablation) | MNIST binary CNN, 256 first-layer kernels · MNIST black-box transfer, FGSM ε=0.3 | accuracy: 70±2% -> 96.4±0.4% | Improved |
| Continue the targeted CW L2 attack from 100 to 1000 iterations | Deterministic binary and FGSM-adversarially-trained binary models · MNIST targeted CW L2 | accuracy: B and B+ reach 19±3% at 1000 iterations | Hurt |
| Use stochastic quantization rather than deterministic binarization under a 1000-step CW L2 attack (ablation) | MNIST binary CNN · MNIST targeted CW L2, 1000 iterations | accuracy: 19±3% -> 38±3% | Improved |
| Use PGD-trained scaled binary rather than PGD-trained full-precision models | MNIST CNN, 256 first-layer kernels · MNIST test, FGSM ε=0.2 | accuracy: 92.9±0.3% -> 93.4±0.3% | Improved |
Closest papers by meaning
- Defensive Quantization: When Efficiency Meets Robustness (ICLR 2019 · similarity 0.89)
- Improved Gradient-Based Adversarial Attacks for Quantized Networks (AAAI 2022 · similarity 0.89)
- Towards Certificated Model Robustness Against Weight Perturbations (AAAI 2020 · similarity 0.88)
- Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations (JMLR 2018 · similarity 0.87)
- Mind the Gap: A Practical Attack on GGUF Quantization (ICML 2025 · similarity 0.85)
- Verifying Properties of Binary Neural Networks Using Sparse Polynomial Optimization (ICLR 2025 · 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.
