Anonymisation of Image Data Using Deep Learning

Sammanfattning: Collecting and storing data is easier and more common than ever before. A lot of this data is personal data, which is problematic to store and use both for ethical reasons and because of legislations. In some scenarios the personal information in the data is not interesting or relevant for the given task but it is still collected and stored as a byproduct of the data collection. In these scenarios it would be much better if one could use anonymised data instead. This report presents a neural network approach for creating an anonymisation filter for image data, specifically for images depicting humans taken from above. A convolutional neural network is used as the filter. It is trained together with a person detector and ReID in order to generate images where it is possible to detect people in the images but impossible to identify them. The training process is similar to that of a generative adversarial network since the goal of the filter was to construct an anonymisation that makes it easy for the detector but hard for the ReID and the aim for the detector and ReID was to become as good as possible. Using the presented method a filter was created which succeeds in anonymising the images whilst almost maintaining the performance of the person detection. Using the metrics AP for the detector and AUC-ROC for the ReID this filter yielded the results 0.864 and 0.540 respectively. However, several of the parameters were very sensitive to changes, yielding results that varied widely when small changes were made, making the network fairly hard to train.