Ketahanan Pembelajaran Mesin terhadap Adversarial examples: Metodologi dan Pertahanan

Abstrak

This paper examines the vulnerability of machine learning models to adversarial examples: inputs that are subtly manipulated to deceive a model into making incorrect predictions. Although deep learning has demonstrated remarkable performance across various tasks, the security of these models remains a significant challenge. This study provides a comprehensive review of various methods for generating adversarial examples, a classification of attack techniques, and corresponding defense strategies, including both active and passive approaches. The findings indicate that a combination of several defense techniques is significantly more effective in enhancing model robustness compared to any single approach. This research is expected to provide a foundation for the development of more secure and reliable machine learning models for critical applications.

Identifier

• OpenAlex: https://openalex.org/W7110000588
• DOI: https://doi.org/10.30998/faktorexacta.v18i2.26078