Summary of experimental settings for Printer Source Attribution (PSA) and License Plate Detection (LPD) tasks
| Aspect | Printer Source Attribution (PSA) | License Plate Detection (LPD) |
|---|---|---|
| Target AI System | Source printer attribution classifier | SSD300-based license plate detector |
| Attack Objective | Induce misclassification of the printer source (Kyocera P5021 CDN) | Cause incorrect localization or suppression of license plate bounding boxes |
| Security Impact | Undermines the reliability of forensic printer authentication | Undermines the reliability of automatic license plate detection systems |
| Dataset Used | VIPPrint dataset | Aggregated dataset compiled from public LPD benchmarks |
| Generative AI Module | Differentiable P&S simulator trained using Pix2Pix and Cycle-GAN | Differentiable P&S simulator trained using CycleGAN |
| EOT Strategy | I-FGSM and C&W attacks combined with P&S simulation and physical transformations | I-FGSM attack combined with P&S simulation and physical transformations |
| Aspect | Printer Source Attribution (PSA) | License Plate Detection (LPD) |
|---|---|---|
| Source printer attribution classifier | SSD300-based license plate detector | |
| Induce misclassification of the printer source (Kyocera P5021 CDN) | Cause incorrect localization or suppression of license plate bounding boxes | |
| Undermines the reliability of forensic printer authentication | Undermines the reliability of automatic license plate detection systems | |
| VIPPrint dataset | Aggregated dataset compiled from public LPD benchmarks | |
| Differentiable P&S simulator trained using Pix2Pix and Cycle-GAN | Differentiable P&S simulator trained using CycleGAN | |
| I-FGSM and C&W attacks combined with P&S simulation and physical transformations | I-FGSM attack combined with P&S simulation and physical transformations |