Issue |
ITM Web Conf.
Volume 70, 2025
2024 2nd International Conference on Data Science, Advanced Algorithm and Intelligent Computing (DAI 2024)
|
|
---|---|---|
Article Number | 02004 | |
Number of page(s) | 8 | |
Section | Machine Learning in Healthcare and Finance | |
DOI | https://doi.org/10.1051/itmconf/20257002004 | |
Published online | 23 January 2025 |
Relieve Adversarial Attacks Based on Multimodal Training
Sydney Smart Technology College, Northeastern University at Qinhuangdao, 066000, Qinhuangdao, China
Corresponding author: 202319117@stu.neuq.edu.cn
This paper explores the role of multimodal training in mitigating the problems caused by adversarial attacks, building on the foundations of deep learning. Deep learning models have reached great success in many areas such as image recognition and natural language processing. But their robustness has always been a concern. However, the emergence of adversarial attacks has exposed shortages of neural networks, forcing people to confront their limitations and further increasing concerns about the security of deep learning models. Adversarial training is an effective defense mechanism that incorporates adversarial samples into the training data, enabling models to better detect and resist attacks. This paper first introduces the principles and types of adversarial attacks, as well as basic concepts and related methods, including Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), DeepFool, and Jacobian Saliency Map Attack (JSMA). The paper then focuses on analyzing the robustness of the multimodal model CLIP based on contrastive learning. Finally, the paper proposes whether audio data can be added to the training samples of the CLIP model to further improve its robustness, and raises related issues and bottlenecks.
© The Authors, published by EDP Sciences, 2025
This is an Open Access article distributed under the terms of the Creative Commons Attribution License 4.0, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.