Black-Box Optimization Based Adaptive Image Anonymization

Arcadi Llanza, Nadiya Shvai, Amir Nakib

Abstract: In the last decade, Convolutional Neural Networks became an industry standard achieving state-of-the-art results for many computer vision tasks. This unprecedented success has been possible due to the use of massive amounts of visual and multimodal data. However, management of these data must comply with the regulations on privacy protection, i.e. personal data should be anonymized. Traditional image anonymization methods such as blurring, masking, pixelating are efficient in the obfuscation of the sensitive data. Still, recent research has indicated that these methods impact in the negative way the performance of computer vision models. Numerous deep learning anonymization methods have been proposed as an alternative, in particular for human face and body anonymization. Unfortunately, the vast majority of these approaches are task-specific and require training. Other methods, although general, rely on full access to the computer vision model (the so-called white-box methods). Here, we propose a novel adaptive image anonymization method that allows one to achieve high concordance of the classification model predictions on the original and anonymized image. It is gradient-free, agnostic to anonymized objects, and to the particular architecture and weights of the computer vision model used. Finally, the proposed method does not require modifications to the computer vision model. The main idea of the approach introduced in this paper is to consider image anonymization as an optimization problem and to solve it using the iFDA metaheuristics algorithm. Experiments conducted on the large-scale benchmark image dataset ImageNet convincingly demonstrate the efficiency of our approach. When applying the proposed adaptive image anonymization method, the class concordance rate obtained was 98.11%, as opposed to 74.47% obtained by traditional anonymization.