Research map / Architectures and efficiency / RNNs and normalization
Radioactive data: tracing through training
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 |
|---|---|---|---|
| Mark 1% of ImageNet training data | Radioactive data with fixed ResNet-18 features · ImageNet, random crop | radioactivity detection significance (log10(p)): -38.0 | Helped |
| Use center crop rather than random crop during training (ablation) | Radioactive data with fixed ResNet-18 features · ImageNet, q=1% | radioactivity detection significance (log10(p)): -38.0 -> <-150 | Helped |
| Train the ResNet-18 from scratch rather than keeping features fixed (ablation) | Radioactive data · ImageNet, q=1% | radioactivity detection significance (log10(p)): -38.0 -> -4.85 | Hurt |
| Add random crop data augmentation (ablation) | Radioactive data trained from scratch · ImageNet, ResNet-18 | radioactivity detection significance (log10(p)): -0.66 -> -4.85 | Helped |
| Add radioactive marking to training data | ResNet-18 trained from scratch · ImageNet, q=1%, random crop | Δacc: -0.1 | Hurt |
| Train with ResNet-50 rather than the marking ResNet-18 (ablation) | Radioactive data · ImageNet, q=2% | radioactivity detection significance (log10(p)): -12.3 vs. -12.63 | Helped |
| Transfer radioactive marks to VGG-16 (ablation) | Radioactive data · ImageNet, q=2% | radioactivity detection significance (log10(p)): -4.49 | Helped |
| Use an ImageNet-pretrained marking network on Places205 (ablation) | Radioactive data · Places205, q=10% | radioactivity detection significance (log10(p)): -3.30 | Helped |
| Use black-box loss comparison instead of white-box inspection (ablation) | Radioactive data · ImageNet, q below 20% | detection | Hurt |
| Distill the marked model before radioactivity testing (ablation) | Radioactive data · ImageNet, q=2% | radioactivity detection significance (log10(p)): -3.07 | Helped |
| Increase the fraction of radioactive training data (ablation) | Radioactive data with ResNet-18 trained from scratch · ImageNet, random crop | radioactivity detection significance (log10(p)): -4.85 -> -12.63 -> -48.8 -> <-150 | Helped |
Closest papers by meaning
- Watermarking Makes Language Models Radioactive (NeurIPS 2024 · similarity 0.84)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.82)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.82)
- Phantom Transfer: Data Poisoning can Survive Data-Level Defences (NeurIPS 2026 · similarity 0.82)
- How Much is a Noisy Image Worth? Data Scaling Laws for Ambient Diffusion. (ICLR 2025 · similarity 0.81)
- Training Data Provenance Verification: Did Your Model Use Synthetic Data from My Generative Model for Training? (CVPR 2025 · similarity 0.81)
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.