Privacy-Preserving Object Recognition with Explainability in Smart Systems

This paper proposes an approach for privacy-preserving object recognition that considers both data privacy protection at different levels and data analysis accuracy while also providing decision explanations. To achieve this, the proposed approach uses multiple degrees of privacy and investigates their impact on the analysis accuracy and the heatmaps generated by the explainability mechanism. The privacy parameter of the privacy-preserving mechanism is regulated for object recognition algorithms and the results’ accuracy is measured accordingly. The methodology uses original images of objects and adds noise to them using Gaussian filter blurring or differential privacy to protect privacy, with three degrees of privacy applied. Object recognition is performed on the original and perturbed images, and the results are compared. To validate our approach, experiments were conducted on 31 categories of real-world object images from the Open Images Dataset V4 using three object recognition models (VGG16, VGG19, and ResNet50). The Grad-CAM mechanism is used to explain the model decisions. The results demonstrate the approach’s effectiveness in protecting data privacy while maintaining data analysis accuracy with the differential privacy mechanism, and providing decision meaningful explanations.