Authors:

Michelle Perez Aguilar

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Securing the machine learning supply chain is crucial due to the increasing reliance on these systems across various sectors and safety-critical applications. Vulnerabilities within the machine learning supply chain can lead to attacks that compromise the entire system, leading to devastating effects such as performance degradation or leakage of private data. In this work, we introduce a novel and covert attack targeting the data loader in machine learning frameworks. By maliciously manipulating the training data sampler, specific class samples can be excluded during the training process, resulting in class imbalance. We demonstrate the feasibility of such an attack and its potential to significantly degrade model performance while remaining difficult to detect. Our implementation reveals that an attacker can undermine model quality and fairness by modifying data loader covertly. Additionally, we discuss security measures that can be implemented to prevent the proposed attack and similar compromises of machine learning systems. Keywords: Cryptographic Protocols; Supply Chain; Life Cycle; Attack; Vulnerable

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

Michelle Perez Aguilar