Evolutionary Algorithm Based Adaptive Image Anonymization

Arcadi Llanza, Nadiya Shvai, Amir Nakib

Abstract: Data privacy laws put restrictions on the processes of image collection, storage, and usage. Commonly sensitive information in an image can be removed by modifying the corresponding areas of image by blurring, masking, or replacement. However, an image anonymized in such a way may result in an unexpected outcome from a neural network performing a computer vision task. In particular, a classification prediction for an anonymized image might significantly differ from a prediction for the original image. In this work, we introduced an adaptive image anonymization method based on evolutionary algorithm that allows to avoid unwanted prediction discrepancy. The proposed approach permits to construct anonymized areas within an image and, it does not require access to the architecture and weights of the neural network that performs classification. Experiments conducted on the ImageNet benchmark dataset with face anonymization convincingly prove the efficiency of the proposed method.