The Impact of Real-World Drone Conditions on the Reliability of Adversarial Patch Attacks*

Session Number

2

Advisor(s)

Pooya Khorrami, Danielle Sullivan, MIT Lincoln Laboratory

Grade Level

A133

Discipline

Computer Science

Start Date

15-4-2026 11:10 AM

End Date

15-4-2026 11:55 AM

Abstract

Computer vision models are algorithms that classify and interpret images, typically through neural networks trained with machine learning. One common application of computer vision is in drones, which are being used for surveillance, defense, and traffic monitoring. Computer vision models can often be vulnerable to adversarial patches attacks. Adversarial patches are created to deceive object detection systems into causing objects to appear, disappear, or being misclassified. This study examines the performance of adversarial patches in real-world conditions by streaming live drone imagery, with and without adversarial patches, through an object detection model. Research investigates how practical variables such as unusual viewing angles, varying lighting conditions, image compression, and motion blur can affect computer vision models in the classification of everyday objects such as cars and people.

Share

COinS
 
Apr 15th, 11:10 AM Apr 15th, 11:55 AM

The Impact of Real-World Drone Conditions on the Reliability of Adversarial Patch Attacks*

Computer vision models are algorithms that classify and interpret images, typically through neural networks trained with machine learning. One common application of computer vision is in drones, which are being used for surveillance, defense, and traffic monitoring. Computer vision models can often be vulnerable to adversarial patches attacks. Adversarial patches are created to deceive object detection systems into causing objects to appear, disappear, or being misclassified. This study examines the performance of adversarial patches in real-world conditions by streaming live drone imagery, with and without adversarial patches, through an object detection model. Research investigates how practical variables such as unusual viewing angles, varying lighting conditions, image compression, and motion blur can affect computer vision models in the classification of everyday objects such as cars and people.