Improving the Robustness of Object Detection and Classification AI models against Adversarial Patch Attacks

📅 2024-03-04
🏛️ arXiv.org
📈 Citations: 3
Influential: 0
📄 PDF
🤖 AI Summary
To address the robustness threat posed by physical-world adversarial patch attacks against the YOLOv5 object detector, this paper proposes a lightweight end-to-end defense framework. The method introduces generative image inpainting as a preprocessing module—its core innovation—semantically reconstructing regions occluded by adversarial patches while preserving both localization accuracy and classification reliability. This is further integrated with model fine-tuning and adversarial sample analysis to establish a robust detection pipeline. Under pixelated physical patch attacks, the framework reduces misclassification rate by over 20%, restoring both classification accuracy and bounding-box localization precision to pre-attack levels. Empirical evaluation demonstrates significant improvements in real-world deployment safety, particularly for safety-critical applications such as traffic sign detection.

Technology Category

Application Category

📝 Abstract
Adversarial patch attacks, crafted to compromise the integrity of Deep Neural Networks (DNNs), significantly impact Artificial Intelligence (AI) systems designed for object detection and classification tasks. The primary purpose of this work is to defend models against real-world physical attacks that target object detection and classification. We analyze attack techniques and propose a robust defense approach. We successfully reduce model confidence by over 20% using adversarial patch attacks that exploit object shape, texture and position. Leveraging the inpainting pre-processing technique, we effectively restore the original confidence levels, demonstrating the importance of robust defenses in mitigating these threats. Following fine-tuning of an AI model for traffic sign classification, we subjected it to a simulated pixelized patch-based physical adversarial attack, resulting in misclassifications. Our inpainting defense approach significantly enhances model resilience, achieving high accuracy and reliable localization despite the adversarial attacks. This contribution advances the resilience and reliability of object detection and classification networks against adversarial challenges, providing a robust foundation for critical applications.
Problem

Research questions and friction points this paper is trying to address.

Defending YOLOv5 against adversarial patch attacks
Evaluating defenses to restore detection confidence
Comparing SAC, Inpainting, and Latent Diffusion Models
Innovation

Methods, ideas, or system contributions that make the work stand out.

Optimized adversarial patches using EigenCAM and grid search
Tested defenses including SAC, Inpainting, Latent Diffusion Models
Latent Diffusion Model restored confidence by 26.61%
🔎 Similar Papers
R
Roie Kazoom
Electrical and Computers Engineering, Ben Gurion University of The Negev, Beer Sheba, Israel
R
Raz Birman
Electrical and Computers Engineering, Ben Gurion University of The Negev, Beer Sheba, Israel
O
O. Hadar
Electrical and Computers Engineering, Ben Gurion University of The Negev, Beer Sheba, Israel