Model hijacking exploitation and mitigation

Chew, Zi Ying (2024) Model hijacking exploitation and mitigation. Final Year Project, UTAR.

Preview

PDF Download (1675Kb) | Preview

Abstract

In today's digital era, the utilization of machine learning has proliferated, making it an invaluable tool across various automotive applications. Machine learning has found its way into numerous facets of automotive engineering and operations. Additionally, machine learning is harnessed in predictive maintenance, where it analyses sensor data from vehicles to forecast potential mechanical issues, thereby optimizing maintenance schedules and minimizing downtime. However, things are two sides. There are vulnerabilities in machine learning that can cause adversarial attack, namely model hijacking attack. In prior research, there are experiment that shows that it is possible to insert trigger into the input to trigger backdoor attack without the user notice. The severity of this problem stems from the attacker's ability to wield this attack method at their discretion, resulting in highly erroneous outcomes. The profound danger lies in the fact that once a malicious actor successfully gains control of a machine learning model, they can manipulate it to generate intentionally misleading results or predictions. Incorrect decisions prompted by a poisoned machine learning model can lead to substantial financial losses, damage to a company's reputation, and even bankruptcy, life and death. In critical applications such as healthcare, autonomous vehicles, and industrial control systems, the reliability of machine learning models is paramount. Thus, in this project, I will focus on the backdoor attack to identify how the attack make use of the loopholes of machine learning. Think from the perspective of an attacker, I propose model that can be backdoor without being realize. It can have face recognition real-time and enter the backdoor mode when physical backdoor is being detected. When the model is in a backdoor mode, it will misclassify the face into the wrong class as attacker intentions. By understanding how the attack works, it can be twisted into a model that gives positives use such as steganography. Thus, in this research, I will develop a backdoor machine learning that can be used for good intention.

Actions (login required)