Privacy-Preserving Machine Learning for Image Data: From Grayscale Single Feature to Color Multi-feature

The increasing demand for privacy protection in machine learning applications has led to the development of various privacy-preserving techniques. Our paper focuses on privacy-preserving machine-learning techniques applied to image classification tasks. Our approach addresses scenarios involving grayscale images with a single sensitive feature and color images with multiple sensitive and public features. We propose a novel approach combining adversarial training and modified loss functions to balance privacy and utility preservation. Specifically, we tackle the data representation problem as (x, s, u), where x represents the raw high-dimensional data, s denotes the sensitive features, and u corresponds to the public or utility features. Through extensive experiments conducted on diverse datasets, we demonstrate the effectiveness of our methods. Our approach safeguards sensitive information while achieving high classification accuracy for public features. The findings from our study provide valuable insights and practical guidance for developing privacy-preserving machine learning models in real-world applications involving image data.