Robust unlearnable examples for data protection in deep learning

Abstract

Object detection is a critical task in computer visionwith diverse applications, including autonomous driving, surveillance, andmedical imaging. Recently, unlearnable examples (UEs) have emerged as a potential solution toprotect sensitive datasets frombeing exploited inmodel training.While substantial research has focused on UEs in image classification tasks, their application toobject detection remains limited. This dissertation investigates the extension of existing UE methods—such asError-Minimizing (EM), Targeted Adversarial Perturbation (TAP), and Linearly Separable Perturbations (LSP)—to object detection models, highlighting the unique challenges that arise due to the dual requirements of localization and classification. Specifically, we explore how these methods could be adapted for use with popular object detection models like YOLOv7, SSD, and EfficientDet. YOLOv7, known for its speed and accuracy, and SSD300, designed for real-time detection, serve as benchmarks for evaluating the impact of UEs on object detection. Additionally, EfficientDet, an efficient and scalable model, is examined for its robustness against such perturbations. Our study aims to bridge the gap between image classification and object detection in the context of UEs and offers insights into how adversarial data poisoning techniques can be applied to prevent unauthorized use of private datasets in training deep learning models.