Motion-Excited Sampler: Video Adversarial Attack with Sparked Prior - Spring 2021

Explore a lecture on video adversarial attacks using the Motion-Excited Sampler technique. Delve into the problem formulation, methodology, and key components such as motion-excited sampling, motion calculation, and gradient estimation. Examine the adversarial optimization process, loss function, and results, including model accuracy and the effectiveness of using motion. Visualize motion vectors and analyze main findings through an ablation study. Gain insights into this innovative approach for generating adversarial examples in videos, presented as part of the CAP6412 course at the University of Central Florida.

Intro
Paper details
Problem Formulation
Method: Overview
Method: Motion-Excited Sampler
Method: Motion Excited Sampling
Method: Motion Calculation
Method: Gradient Estimation - Formulation
Method: Gradient Estimation - Big Idea
Method: Gradient Estimation - Algorithm
(Method) Adversarial Optimization
(Method) loss function
(Results) Model Accuracy
(Results) Effectiveness of using motion
(Results) Visualizing Motion Vectors
(Results) Main Results
(Results) Ablation study
Summary
Against