Research map / Adversarial robustness and security / Certified and empirical robustness
Efficient Exact Verification of Binarized 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 BinMask sparsifier instead of ternary quantization (ablation) | conv-small BNN, MNIST eps=0.1 · MNIST | mean solve time (s): 756.288 -> 0.002 | Helped |
| use BinMask instead of ternary quantization (ablation) | conv-small BNN · CIFAR10 eps=2/255 | mean solve time (s): 0.267 -> 0.003 | Helped |
| use BinMask instead of ternary quantization (ablation) | conv-small BNN · MNIST / CIFAR10 | natural test accuracy: 97.59% -> 97.35% (MNIST); 54.78% -> 55.22% (CIFAR10) | No clear effect |
| add cardinality bound decay loss (eta=1e-4) (ablation) | conv-large BNN adversarially trained · MNIST eps=0.3 | mean solve time (s): 2442.698 -> 0.009 | Helped |
| add cardinality bound decay loss (ablation) | conv-large BNN · MNIST eps=0.3 | natural test accuracy: 98.88% -> 96.97% | Hurt |
| add cardinality bound decay loss (ablation) | conv-large BNN · CIFAR10 eps=8/255 | PGD accuracy: 15.32% -> 27.06% | Helped |
| add cardinality bound decay loss (ablation) | conv-large BNN · CIFAR10 eps=8/255 | natural test accuracy: 53.91% -> 40.80% | Hurt |
| use tanh gradient cancelling instead of hard tanh (ablation) | conv-large BNN adversarial training · CIFAR10 eps=8/255 | verifiable accuracy: 11.13% -> 14.70% | Helped |
| use adaptive gradient cancelling instead of tanh (ablation) | conv-large BNN · CIFAR10 eps=8/255 | verifiable accuracy: 14.70% -> 20.08% | Helped |
| add verifier-generated adversarial examples in last ten epochs (ablation) | conv-large BNN adaptive · CIFAR10 eps=8/255 | verifiable accuracy: 20.08% -> 22.55% | Helped |
| use tanh instead of hard tanh gradient in PGD attack (ablation) | conv-small network · CIFAR10 | attack success rate: 19.82% -> 23.25% | Helped |
| use BNN with EEV instead of exact verification of real-valued nets | conv-small/large · CIFAR10 eps=2/255 | verifiable accuracy: 45.93% -> 26.13% (small) | Hurt |
Closest papers by meaning
- Verifying Properties of Binary Neural Networks Using Sparse Polynomial Optimization (ICLR 2025 · similarity 0.90)
- VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees (ICML 2024 · similarity 0.87)
- Certified but Private: Scalable Zero-Knowledge Proofs for the Formal Verification of Neural Networks (NeurIPS 2026 · similarity 0.87)
- Tightening Robustness Verification of MaxPool-based Neural Networks via Minimizing the Over-Approximation Zone (CVPR 2025 · similarity 0.86)
- Training Verification-Friendly Neural Networks via Neuron Behavior Consistency (AAAI 2025 · similarity 0.86)
- Towards General Robustness Verification of MaxPool-based Convolutional Neural Networks via Tightening Linear Approximation (CVPR 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.