Validation report 003: YOLOv5-license-plate exploration with SHAP

This project investigates the robustness of a YOLOv5-based model (Ultralytics 2021), finetuned for license plate detection, against adversarial attacks using a variation of DPatch Liu et al. (2019). Additionally, we explore the interpretability of the model’s predictions through SHAP (Lundberg and Lee 2017).

This study examines the robustness and interpretability of a fine-tuned YOLOv5 model for license plate detection using DPatch adversarial attacks and SHAP analysis.

DPatch experiments revealed limitations in perturbing the model, with reduced detection rates. Discrepancies from the original DPatch paper were attributed to specific fine-tuning for license plate detection and potential model advancements.

SHAP analysis provided insights into the model’s decision-making, highlighting the local effects of DPatch and the model’s focus on specific regions, such as the license plate. Numbers from Tables 1-3 align with our observations and further solidify our understanding on DPatch attacks.

Link to original publication with a model: yolov5m-license-plate

Recommended citation: Robert Laskowski, Szymon Sadkowski. (2024). "Interpreting License Plate Detection Model: A SHAP-based Analysis and Adversarial Attack Exploration." Github: ModelOriented/CVE-AI.
Download Paper