Logo image
Adversarial attacks against datasets and algorithms
Thesis   Open access

Adversarial attacks against datasets and algorithms

Rudzani Nndwa
Master of Artificial Intelligence, University of Johannesburg
2024
Handle:
https://hdl.handle.net/10210/519311

Abstract

The study reveals that both CNNs and SVMs are significantly sensitive to adversarial manipulation, leading to drastic reductions in accuracy. Through the application of adversarial training, the study identifies a trade-off between enhancing robustness against adversarial examples and maintaining performance on clean data. The findings emphasise the importance of optimizing model architectures, hyperparameters, and training strategies to mitigate vulnerabilities. Recommendations for future studies include developing more sophisticated adversarial training methods, exploring alternative machine learning models, and tailoring defence strategies to specific realworld applications. This study contributes to the field of adversarial machine learning by providing empirical evidence and insights that inform the development of more secure and reliable machine learning systems.
pdf
Nndwa R 215026446 (4)1.06 MBDownloadView
Open Access

Metrics

1 Record Views

Details

Logo image