Revisit linear transformation for image privacy in machine learning

Linear transformation (LT) plays an important role in machine learning and data science, from data representation, to machine-learning (ML) model training. On the other hand, a well known machine learning model for visual tasks, CNN, is can be interpreted as a LT system into multiple layers which are connected by non-linear units, i.e., activation functions. In this paper, we revisit LT as a highly-efficient approach to encrypt images for ML applications, meanwhile the ML performance over encrypted images is comparable to ML over original images. The preliminary experimental results demonstrate that LT is able to successfully high the visual contents but has little effect on CNN model performance. In addition, LT is suitable for different computer vision tasks, such as image classification and object detection, making LT a good candidate for privacy-preserving machine learning as a service (MLaaS) for visual tasks.