Attack on training effort of deep learning

Abstract

The objective of this project is to develop an attack to hinder the tracking results of state-of-the- art Visual Object Trackers. After code development and testing, an evaluation will be done to assess the performance of the attack and to draw conclusions. Visual Object Tracking is a relatively new technology with increasing usage in modern systems. As Visual Object Trackers are built using deep learning models, it is inherently prone to the same vulnerabilities which give rise to the need to properly secure such systems. This project aims to attack Visual Object Trackers through the means of data poisoning with adversarial examples. An attack script was developed to utilise consecutive frames from a video to synthesize motion blurred images which are then used to poison the dataset that the object tracker is working on. The mechanisms implemented and inner workings were detailed, and an evaluation was drawn on the performance of the developed attack script. The attack script performed to expectation and was successful in achieving the goals set out for this project. This allows for further research to explore similar attacks in detail to devise appropriate protection/counter mechanisms against them.