基于模态一致性的协同训练多模态学习一种隐私保护方法.PDFVIP

基于模态一致性的协同训练多模态学习：一种隐私保护方法 苗 园 詹德川 （南京大学计算机软件新技术国家重点实验室 南京 210023 ） 摘 要 复杂的数据对象通常是由来自不同信息源或提取方式的多模态特征描述的。因此，多模态学习引起了研究者的广泛 关注。通过利用多个模态信息之间的关系，多模态学习能够很好地提升学习效果。 在一些真实应用场景中，不同模态的数 据特征可能来自于不同的私有数据源，要求在学习模型训练和测试时数据源之间的信息不能共享，也即在利用多模态之间的 关系的同时保护数据隐私。现有的多模态学习方法中，子空间方法和先融合方法不适用于这种场景，因为这两类方法需要使 用所有的原始数据特征；后融合方法基于模态个体分类器的预测结果的，但无法利用模态之间的关系来提升个体分类器的性 能；协同训练方法难以处理包含多个模态的数据，因此也不能适用于该场景。本文定义了一种新的模态一致性指标，在协同 训练框架下通过利用所有其它模态个体分类器对未标记数据的预测结果的后融合来增大某个模态的训练集，提出了一种新 颖的保护数据特征隐私的多模态学习方法。多个真实数据集上的实验结果显示了该算法的有效性。 关键词 机器学习，半监督，多模态，协同训练，隐私保护 中图法分类号 TP391.4 文献标识码 A View Consistency based Co-training style Multi-View Learning: A Privacy-Preserving Approach MIAO Yuan ZHAN De-chuan (National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China) Abstract Complex media objects are often described by multi-view feature groups collected from diverse domains or information channels. Multi-view learning, which attempts to exploit the relationship among multiple views to improve learning performance, has drawn extensive attention. It is noteworthy that in some real-world applications, features of different views may come from different private data repositories, and thus, it is desired to exploit view relationship with data privacy preserved simultaneously, i.e. not to share data features with other data repositories when training and testing the model. Existing multi-view learning approaches such as subspace methods and pre-fusion methods are not applicable in this scenario because they need to access the whole features, whereas late-fusion approaches could not exploit information from other views to improve the individual view-specific learners. Co-training methods can hardly deal with multiple views. Under the framework of co-training, by using view consistency defined on all th