Ensuring Data Privacy in Deep Learning through Compressive Learning

近年来，深度学习算法在多个参与者参与算法训练过程的协作环境中得到了广泛应用。例如，用户可以提交图像，以便集体使用，以训练用于图像分类的机器学习模型。协作学习的一个主要问题是保护参与者的隐私；这可能是指隐藏他们的身份或他们提供的数据。在许多情况下，我们希望确保在共享模型或模型更新时，数据不能直接与特定的个人关联。在本提案中，我们将开发使用压缩学习方法来学习具有差分隐私（DP）保证的深度学习模型的新方法。我们将使用一个涉及人工智能辅助视频内容审核的数据集，其中使用基于深度学习的分割和分类模型来识别视频中的明确图像内容，并为此类内容建议适当的级别（中度，严重等）。本文的研究避免了现有方法的缺点，实现了更低的计算成本，适用于更一般的数据和分析场景。因此，它将消除目前将私人深度学习大规模应用于人工智能辅助视频内容审核的计算障碍。能够保证个人训练图像的内容是私有的，将有助于最大限度地减少内容泄露的风险，从而使较年轻的年龄组接触到不合适的泄露内容。

Recent years have seen the wide application of deep learning algorithms in a collaborative setting where multiple participants contribute to the training process of the algorithm. For example, users may submit images to be used collectively to train a machine learning model for image classification. A major concern of collaborative learning is protecting privacy of the participants; this could refer to concealing either their identity or the data they provide. In many cases, we want to make sure that data cannot be directly associated with a specific individual when the model or updates to the model are shared.In this proposal, we will develop new methods to learn deep learning models with differential privacy (DP) guarantee using compressive learning approaches. We will work with a dataset that concerns AI-assisted video content moderation, in which deep-learning based segmentation and classification models have been used to identify explicit image content in the videos and to suggest appropriate levels (moderate, severe etc) for such content. The proposed research avoid the drawbacks of current approaches as well as achieve lower computational cost and be applicable in more general data and analysis scenarios. it will thus remove current computational barriers of applying private deep learning for AI-assisted video content moderation at scale. Being able to guarantee that the content of individual training images are private will help minimize the risk that content is leaked and thus that younger age groups are exposed to unsuitable leaked content.