PIL generates unlearnable examples using lightweight linear proxies instead of DNNs, slashing compute costs. It exploits the insight that perturbations induce linear behavior in deep models. Offers comparable protection with far superior efficiency on CIFAR/ImageNet.
Key Points
- 1.Replaces DNN proxies with linear models for PGD optimization
- 2.Key insight: unlearnable samples boost model linearity (FGSM metric)
- 3.15+ GPU hours (REM) → minutes; scales to high-res images
- 4.Open-source code protects data privacy without heavy compute
Impact Analysis
Democratizes unlearnable examples for photographers/users, enabling practical data protection against model scraping at low cost.
Technical Details
Optimizes perturbations on linearized proxies; transferability holds due to induced linearity across methods like EM/REM/TAP.
