Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Reprogramming of Neural Networks
Gamaleldin F. Elsayed, Ian Goodfellow, Jascha Sohl-Dickstein
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 |
|---|---|---|---|
| Add a shared learned input program to repurpose ImageNet classifiers for square counting | ImageNet classification models · Counting squares | accuracy: 0.9993; 0.9999; 0.9994; 0.9763; 0.9843; 0.9966 | Improved |
| Add a shared learned input program to repurpose ImageNet classifiers for digit classification | ImageNet classification models · MNIST test | accuracy: 0.9753; 0.9646; 0.9744; 0.9534; 0.9664; 0.9496 | Improved |
| Add a shared learned input program to repurpose ImageNet classifiers for object classification | ImageNet classification models · CIFAR-10 test | accuracy: 0.6911; 0.6683; 0.6719; 0.6210; 0.6301; 0.5858 | Improved |
| Add a learned program for shuffled rather than normally structured MNIST digits | ImageNet classification models · Shuffled MNIST test | accuracy: 0.9709; 0.9715; 0.9683; 0.9691; 0.9678; 0.9717 | Improved |
| Use adversarially trained rather than standard Inception V3 weights (ablation) | Adversarial reprogramming for MNIST · MNIST test | accuracy: 0.9753 -> 0.9752 | No clear effect |
| Initialize the target network randomly instead of using ImageNet-trained weights (ablation) | Adversarial reprogramming for MNIST · MNIST test | accuracy: 0.9753 -> 0.4539; 0.9646 -> 0.1861; 0.9744 -> 0.1135; 0.9534 -> 0.1032; 0.9664 -> 0.1756; 0.9496 -> 0.9325 | Hurt |
| Shuffle the pixels of adversarial MNIST inputs (ablation) | Adversarial reprogramming for MNIST · Shuffled MNIST test | accuracy: 0.9709; 0.9715; 0.9683; 0.9691; 0.9678; 0.9717 | No clear effect |
| Limit the size of the adversarial program (ablation) | Inception V3 reprogrammed for MNIST · MNIST classification | accuracy | Hurt |
| Limit the perturbation scale to make the program nearly imperceptible (ablation) | Inception V3 reprogrammed for MNIST · MNIST classification | accuracy | Hurt |
| Hide shuffled MNIST data and the program inside a normal ImageNet image (ablation) | Inception V3 reprogrammed for MNIST · MNIST classification | accuracy | Hurt |
Closest papers by meaning
- Adversarial Reprogramming Revisited (NeurIPS 2022 · similarity 0.89)
- Endowing Visual Reprogramming with Adversarial Robustness (ICLR 2025 · similarity 0.87)
- Generative Adversarial Perturbations (CVPR 2018 · similarity 0.87)
- Deep Defense: Training DNNs with Improved Adversarial Robustness (NeurIPS 2018 · similarity 0.87)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · similarity 0.86)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 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.
