Study of Adversarial Machine Learning with Infrared Examples for Surveillance Applications

Demarcus Edwards,Danda B Rawat

doi:10.3390/electronics9081284

Abstract

Adversarial examples are theorized to exist for every type of neural network application. Adversarial examples have been proven to exist in neural networks for visual-spectrum applications and that they are highly transferable between such neural network applications. In this paper, we study the existence of adversarial examples for Infrared neural networks that are applicable to military and surveillance applications. This paper specifically studies the effectiveness of adversarial attacks against neural networks trained on simulated Infrared imagery and the effectiveness of adversarial training. Our research demonstrates the effectiveness of adversarial attacks on neural networks trained on Infrared imagery, something that hasn’t been shown in prior works. Our research shows that an increase in accuracy was shown in both adversarial and unperturbed Infrared images after adversarial training. Adversarial training optimized for the L∞ norm leads to an increase in performance against both adversarial and non-adversarial targets.

Highlights

Surveillance is a common application of Infrared cameras and the detection/recognition tasks in the electromagnetic band are common in military and surveillance applications
Information that is shown in the Visual such as reflectively, luminescence, color, and texture, is lost either partially or entirely when looking at the same object in the Infrared spectrum. This is the motivation of this paper showing that exploiting the gradients of an image can be an effective attack against computer vision models for surveillance trained on Infrared imagery and that the same defenses being deployed for models trained for the visual spectrum can be deployed models trained for the Infrared spectrum
The IR CNN improved by 13% to 60% accurate on the test set while being 46% accurate when testing on adversarial images, a full 35% increase in accuracy against adversarial images

Summary

Introduction

Surveillance is a common application of Infrared cameras and the detection/recognition tasks in the electromagnetic band are common in military and surveillance applications. Real-time operations such as tracking work through running an object recognition algorithm on frames of a video. Information that is shown in the Visual such as reflectively, luminescence, color, and texture, is lost either partially or entirely when looking at the same object in the Infrared spectrum. This is the motivation of this paper showing that exploiting the gradients of an image can be an effective attack against computer vision models for surveillance trained on Infrared imagery and that the same defenses being deployed for models trained for the visual spectrum can be deployed models trained for the Infrared spectrum

Objectives

Results

Conclusion