多模态图像智能目标识别对抗攻击

Intelligent multi-modal image recognition for adversarial attacks

基于深度学习模型的新一代智能化多模态(可见光/红外/雷达)图像识别系统已逐步在航空航天情报侦察、人机交互增强作战系统、无人作战平台自动图像目标识别以及多模复合图像末制导等多个军事场景中得到广泛应用。然而,由于深度神经网络模型在理论上存在不完备性和对抗脆弱性、多模态图像目标识别深度网络结构设计与优化在工程上存在迁移性等因素,使得现有识别系统在鲁棒准确性方面评估不足,给系统在未来战场复杂对抗场景中的广泛部署带来极大的安全隐患。为此,本文通过研究多模态图像智能目标识别系统军事场景应用的风险模型,分析系统存在的潜在攻击面,开展基于深度神经网络的多模态图像识别对抗样本攻击技术和对抗鲁棒准确性评估等关键技术研究,以期提升系统在复杂电磁环境条件下的鲁棒性和准确性。

The new generation of intelligent multi-modal image recognition systems based on deep computing models is widely used in military scenarios,such as intelligent aerospace reconnaissance,human-computer interaction-enhanced combat systems,automatic recognition of unmanned combat platforms through images,and multi-modal image-based terminal guidance.However,owing to the theoretical incompleteness of the deep computing model,the accuracy of the networks used for image recognition in prevalent multi-modal image-based target recognition systems has not been adequately assessed.Thus,the widespread deployment of such systems in complex confrontational military scenarios in the future poses significant security risks.It is therefore important to theoretically establish a risk model for the application of intelligent image recognition systems to military scenarios.Considerable work is also needed on potential attack surfaces,the development of image recognition countermeasure attack technology based on the deep neural network,and accuracy assessment of the models.The ultimate goal is to improve the robustness and accuracy of the system in testing military scenarios.