Quantum-Resilient Federated Learning with Differential Privacy for Secure Content-Based Image Retrieval in Cloud Environments

Abstract

Content based image retrieval (CBIR) in the cloud needs a model that describes images well, but the images often cannot be pooled, federated model updates can leak them, and the public key cryptography protecting those updates will not survive a quantum computer. This paper presents a retrieval framework that meets all three concerns in one round. Each client passes its images through a frozen public ResNet-50 and releases four summary statistics of the features with Gaussian noise for record level differential privacy. Secure aggregation with masks agreed through ML-KEM-768 sums the statistics, and the server builds a regularised linear discriminant projection from the noisy sum. On Caltech-101, CIFAR-10 and Tiny ImageNet the non-private version raises mean average precision from 79.5, 45.2 and 24.7 percent to 90.4, 78.8 and 43.4 percent, respectively. With a privacy budget of epsilon equal to 8 the cost is 7.2, 2.0 and 1.7 points, and it falls as the data per class grows. The result does not depend on how the data is split between clients, a membership inference attack stays near chance, and each client uploads 1.1 to 1.9 MB once, 91 to 93 percent less than 30 rounds of federated averaging. A single private round is therefore enough to build a useful shared retrieval space without pooling the images or exposing the model updates to a future quantum attacker