自然语言处理深度模型的对抗攻防关键算法研究

自然语言处理基于深度学习技术迎来了新的突破，但其模型的鲁棒性问题也逐渐引起了国内外学者的关注。文本对抗样本是指对文本做精心设计的扰动，不影响人类阅读却能使模型产生错误输出的一类样本。该方向近两年刚刚兴起，尚有不少空白。由于文本的离散性质，无法直接借鉴图像对抗攻防这一较为成熟领域的方法。本项目拟从深度学习对抗性攻击的生成机理、对抗攻击与防御和攻防实证几个层面展开对自然语言处理深度模型的对抗攻防基础理论和关键算法的研究，基于强化学习、遗传算法、快速梯度投影、句法结构转换等方法设计高效、快速的对抗攻击方法，基于对抗训练、梯度投影惩罚等方法设计可有效抵御现有攻击且具有一定能力抵御未知攻击的、高可靠性深度模型系统，并将这些基础理论与关键技术用于文本分类、机器翻译等场景的对抗攻防实证研究。本项目具有一定的前瞻性且有重要的应用需求，将为相关核心应用提供理论和技术支撑，产生聚焦国际前沿的原创性研究成果。

Natural Language Processing (NLP) has made great progress with the rapid development of deep learning techniques, and the robustness of deep NLP models has attracted more attention recently. A text adversarial example is a text generated by adding well-designed perturbation to the original text that does not influence human reading but leads to wrong output. This is a new research direction started in recent two years, and there are many problems to be explored. Due to the discrete property of text, techniques developed for image adversarial attack or defense are hard to be implemented for NLP.. In this project, we will focus on the basic theory and key algorithms for defense and attack on various NLP deep learning models, including the intrinsic reason, adversarial attack and defense methods and applications. We will design highly effective adversarial attack methods based on reinforce learning, genetic algorithm, fast gradient projection method, syntactic structure transformation. Also, based on adversarial training and gradient projection penalty, we will design novel methods to effectively resist existing attacks, and wish the designed models have a high reliability to future possible attacks. The developed theories and methods will be applied to the sceneries of text classification and machine translation. This is a forward-looking project that has important applications. It will provide theoretical as well as technical support for robust deep learning applications, and it is expected to yield world class, fundamental results.