2026 Volume 34 Pages 715-730
We propose a privacy-preserving machine learning framework tailored for image data, called PPML with Utility Across Trust Boundaries (UAT). This framework supports both trusted and untrusted testing environments, where the model is trained on perturbed data to ensure privacy against an untrusted learning server. In untrusted testing, test data must also be perturbed, requiring the model to maintain high utility despite the added noise. In trusted testing, test data remain unperturbed but must be adapted to a model trained on noisy inputs in order to achieve even higher utility. To address these challenges, we introduce image-specific local dimension reduction at the user side, extending WA, and propose CWALDP, which clusters pixels and labels based on image characteristics and allocates the privacy budget per cluster. This enables controllable utility under a fixed total privacy budget. Experiments on the MNIST and FMNIST datasets using Support Vector Machines, Random Forests, and Convolutional Neural Networks demonstrate that our PPML with UAT achieves high accuracy even when users locally perturb their images under LDP. We also provide a theoretical evaluation by quantifying the picture difference between images anonymized by CWALDP and the original images.