在视觉-语言预训练（VLP）模型领域，南开大学与南洋理工大学的研究团队共同推出了一项创新策略，旨在增强VLP模型对抗迁移性，特别是针对图像理解任务的准确性。这一研究聚焦于对抗攻击中的对抗轨迹，旨在解决现有研究中对抗样本高度依赖代理模型生成，且存在代理模型过拟合风险的问题。这一发现有望为VLP模型的开发与应用带来突破性进展，提高其在实际场景中的表现。
研究团队首次引入了对抗轨迹交集区域的概念，这一区域由干净样本、当前对抗样本以及上一步对抗样本共同构成。通过这一策略，研究者能够更精确地识别并分析对抗样本的生成过程，从而有效减少代理模型的过拟合风险，提升VLP模型在对抗攻击中的稳健性和鲁棒性。
南开大学的大四学生高森森作为该研究的第一作者，其实习导师郭青研究员（郭青主页：https://tsingqguo.github.io）对研究结果给予了高度评价。郭青表示：“通过引入对抗轨迹交集区域的概念，我们能够更深入地理解对抗样本生成的动态过程，从而针对性地优化VLP模型的训练策略，显著提高其在图像理解任务中的准确性。”
此外，南洋理工大学的加小俊博后研究员（加小俊主页：https://jiaxiaojunqaq.github.io）作为共同第一作者和共同通讯作者，对这一研究成果的贡献表示认可：“我们的研究不仅揭示了对抗攻击中的新视角，更重要的是，它提供了一套有效的策略，为VLP模型的开发者和研究者在对抗攻击防御方面提供了新的思路和工具。”
此研究不仅为学术界提供了宝贵的见解，而且为VLP模型在实际应用中的安全性与可靠性提供了坚实的理论基础。未来，随着更多研究的深入，这一创新策略有望在自动驾驶、医疗影像分析、社交媒体内容审核等领域的VLP应用中发挥重要作用，推动人工智能技术的健康发展。
通过此次合作，南开大学与南洋理工大学的研究团队展现了在人工智能领域协同创新的能力，为解决VLP模型对抗攻击中的关键挑战提供了新的解决方案，进一步推动了全球AI学术研究与产业应用的前沿发展。
英语如下：
News Title: “Innovative Strategies to Enhance VLP Model’s Robustness Against Adversarial Attacks: New Research from Nankai-Nanyang Tech Report for ECCV 2024”
Keywords: Adversarial Attacks, VLP Model, Strategy Enhancement
News Content: Title: Nankai University and Nanyang Technological University Collaborate to Introduce a Novel Strategy for Improving the Accuracy of VLP Models Against Adversarial Attacks
In the domain of Visual-Language Pre-Training (VLP) models, a research team from Nankai University and Nanyang Technological University has jointly introduced an innovative strategy aimed at enhancing the VLP model’s resistance to adversarial attacks, particularly in the accuracy of image understanding tasks. This research focuses on the adversarial trajectories in the context of adversarial attacks, addressing the issue of high reliance on proxy models for generating adversarial samples and the risk of overfitting in these proxy models. This discovery is expected to bring a breakthrough in the development and application of VLP models, improving their performance in real-world scenarios.
For the first time, the research team has introduced the concept of the intersection area of adversarial trajectories, which is composed of clean samples, the current adversarial sample, and the adversarial sample from the previous step. By implementing this strategy, the researchers are able to more precisely identify and analyze the generation process of adversarial samples, effectively reducing the risk of overfitting in proxy models and enhancing the robustness of the VLP model against adversarial attacks.
Gaosensen, a fourth-year student from Nankai University, has taken the lead as the first author of this study, with Dr. Qing Guo (Guo Qing’s homepage: ) as his internship supervisor, who has highly praised the research results. Dr. Guo states, “By introducing the concept of the intersection area of adversarial trajectories, we are able to delve deeper into the dynamics of adversarial sample generation, thereby optimizing the training strategies for VLP models in a targeted manner, significantly improving their accuracy in image understanding tasks.”
Jiaxiaojun, a post-doctoral researcher at Nanyang Technological University (Jiaxiaojun’s homepage: ) as a co-first author and co-corresponding author, acknowledges the contribution of this research result, stating, “Our research not only sheds light on new perspectives in the context of adversarial attacks, but more importantly, it provides an effective strategy for developers and researchers of VLP models to enhance their defenses against adversarial attacks.”
This study not only provides valuable insights for the academic community but also lays a solid theoretical foundation for the security and reliability of VLP models in practical applications. Looking ahead, as more research progresses, this innovative strategy is expected to play a significant role in the application of VLP models in fields such as autonomous driving, medical image analysis, and social media content moderation, driving the advancement of AI technology.
Through this collaboration, the research teams from Nankai University and Nanyang Technological University have demonstrated their capability for collaborative innovation in the field of artificial intelligence, providing new solutions to critical challenges in VLP models’ resistance to adversarial attacks, and further propelling the frontiers of global AI academic research and industrial applications.
【来源】https://www.jiqizhixin.com/articles/2024-07-22-6
Views: 0