A Visual Privacy Protection Method For Visual Crowd Sensing Applications

Visual Crowd Sensing(VCS)is a sensing paradigm of mobile crowdsensing(MCS),which can leverage built-in cameras of smart devices to attain informative and comprehensive sensing of interesting targets,it is a very promising technology.The key process of VCS applications is to coordinate users to collect visual data,i.e.images or videos,and upload them to the cloud server.However,during taking a photo or recording a video,the privacy information of others is captured inadvertently,which resulting in infringement of the privacy for others.This problem will seriously hinder the development of VCS technology,so we need a visual privacy protection method to solve this problem.Until now there is no visual privacy protection method for VCS applications.Fortunately,our work makes up for this deficiency.This paper proposes a new VCS application framework that can protect effectively the visual privacy security in VCS applications.Compared with traditional VCS framework,the proposed visual privacy protection mechanism realizes the privacy-preserving of VCS applications by detecting the privacy-object in pictures or videos in real time and then eliminating them.To achieve privacy-object detecting,a deep learning based object detection model is designed,which can be trained with privacy-object annotation image data to detect specific privacy-object.Moreover,as privacy-object targets increase,privacy-object detection model is required to update frequently.If we use traditional training method to update the model,we need to re-train the model with both new training data and previous training data.But due to privacy protection constraint,the previous privacy-object annotation image data may not be retained for a long time,which may cause the traditional training method to fail.To solve this problem,we innovatively propose an incremental learning method that is suitable for object detection technology.The method only uses the new training data to update model on the basis of the existing privacy-object detection model without requiring the previous training data.It effectively overcomes the drawbacks of traditional training methods and improves the efficiency of model updating procedure.Finally,we complete privacy protection of VCS through fuzzy processing of the detected privacy-object position.We have conducted extensive experimental studies on public object detection data sets and simulated privacy data sets.The experimental results show the effectiveness of the privacy-object detection algorithm and incremental learning method respectively,and verifies the feasibility of the visual privacy protection framework proposed in this paper.