Research map / Adversarial robustness and security / Certified and empirical robustness
Random and Adversarial Bit Error Robustness: Energy-Efficient and Secure DNN Accelerators
David Stutz, Nandhini Chandramoorthy, Matthias Hein, Bernt Schiele
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 per-layer instead of global quantization (ablation) | Normal quantization · CIFAR10 8-bit p=0.5% | robust test error: 90.71 -> 24.76 | Improved |
| use asymmetric quantization into signed integers (ablation) | per-layer quantization · CIFAR10 8-bit p=0.5% | robust test error: 24.76 -> 40.78 | Hurt |
| quantize into unsigned integers (ablation) | asymmetric quantization · CIFAR10 8-bit p=0.5% | robust test error: 40.78 -> 17.00 | Improved |
| use proper rounding instead of integer conversion (ablation) | RQuant · CIFAR10 8-bit p=0.5% | robust test error: 17.00 -> 11.28 | Improved |
| use proper rounding at 4 bit (ablation) | Clipping[0.1]+RQuant · CIFAR10 4-bit p=0.5% | robust test error: 90.36 -> 7.71 | Improved |
| add weight clipping wmax=0.1 | RQuant · CIFAR10 8-bit p=1.5% | robust test error: 68.65 -> 12.22 | Improved |
| use per-layer instead of global weight clipping (ablation) | Clipping · CIFAR10 8-bit p=1.5% | robust test error: 12.22 -> 8.14 | Improved |
| add label smoothing to weight clipping (ablation) | Clipping[0.15] · CIFAR10 8-bit p=1% | robust test error: 13.08 -> 29.40 | Hurt |
| use batch norm instead of group norm (ablation) | Clipping[0.1] · CIFAR10 8-bit p=1% | robust test error: 8.93 -> 18.32 | Hurt |
| tighten clipping to wmax=0.025 (ablation) | Clipping · CIFAR10 8-bit p=1% | robust test error: 7.18 -> 8.18 | Hurt |
| add random bit error training to global clipping (ablation) | Clipping[0.1] · CIFAR10 8-bit p=1.5% | robust test error: 12.22 -> 8.65 | Improved |
| add random bit error training to per-layer clipping (ablation) | PLClipping[0.25] · CIFAR10 8-bit p=1.5% | robust test error: 8.14 -> 7.02 | Improved |
Closest papers by meaning
- Defensive Quantization: When Efficiency Meets Robustness (ICLR 2019 · similarity 0.87)
- Double-Win Quant: Aggressively Winning Robustness of Quantized Deep Neural Networks via Random Precision Training and Inference (ICML 2021 · similarity 0.87)
- Improving Robustness Against Stealthy Weight Bit-Flip Attacks by Output Code Matching (CVPR 2022 · similarity 0.86)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · similarity 0.85)
- NeuralFuse: Learning to Recover the Accuracy of Access-Limited Neural Network Inference in Low-Voltage Regimes (NeurIPS 2024 · similarity 0.85)
- Enhancing Adversarial Robustness in SNNs with Sparse Gradients (ICML 2024 · 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.
