基于高阶TV正则化低秩矩阵重构的深度神经网络鲁棒性增强研究

随着深度学习研究的不断深入，其鲁棒性不足问题受到研究者的广泛关注。即当输入样本受到一定的对抗干扰时，原先得到的深度神经网络模型将无法对该样本进行有效识别。因此，若忽视其潜在的鲁棒性隐患，深度学习的应用，如人脸识别支付等，不仅会让攻击者乘虚而入，而且会给社会造成不可估量的损失。本项目拟结合现有的防御机制，充分利用低秩矩阵重构的图像结构探索能力及高阶TV范数正则化的区域光滑增强优势，探索基于高阶TV正则化低秩矩阵重构的样本预处理机制，提升对抗噪声消除效果，增强深度神经网络的鲁棒性。具体研究内容包括：（1）构建高阶TV正则化低秩矩阵重构模型及其快速求解算法；（2）设计基于关键点检测及低秩矩阵重构的噪声消除防御方法；（3）结合GAN网络，搭建基于高阶TV正则化低秩矩阵重构的对抗噪声消除网络模型。本项目的研究，不仅有利于拓展低秩矩阵理论及其应用范围，而且为解决深度神经网络的鲁棒性提供新思路。

With the deepening of deep learning research, its problem of robustness has attracted widespread attention from researchers. When the input samples are corrupted by certain adversarial disturbance, the original deep neural network model will not achieve the expected recognition effect. Therefore, if the potential robust risks are neglected, the application of deep learning, such as payment based on face recognition, will not only allow malicious attackers to take advantage of it, but also cause immeasurable losses to society. This project intends to combine the existing defensive framework, make full use of the low-rank matrix reconstruction image structure exploration ability and high-order TV norm regularization of regional smooth enhancement advantages, develop images pre-processing mechanism based on high-order TV regularization low-rank matrix reconstruction, so as to better improve the ability to eliminate noise and enhance the robustness of deep neural networks. The specific research contents include: (1) Constructing low-rank matrix reconstruction models with higher-order TV regularization and their fast solving algorithms; (2) Designing a de-noising defense method based on key point detection and low-rank matrix reconstruction; (3) Combing GAN network, build an adversarial noise removing optimization model based on low-rank matrix reconstruction with higher-order TV regularization. The research of this project not only expands the theory of low-rank matrix and its application scope, but also provides new ideas for solving the stability of deep neural networks.