HE-CloudML: a privacy-preserving framework for secure machine learning inference over encrypted cloud data using homomorphic encryption

Abstract The widespread adoption of cloud-based Machine Learning as a Service (MLaaS) exposes sensitive user data to critical privacy risks during inference, as plaintext data must typically be processed by untrusted cloud servers. This paper presents HE-CloudML, a unified privacy-preserving framework for secure deep neural network (DNN) inference over encrypted cloud data using Homomorphic Encryption (HE). HE-CloudML is architected as a three-tier system comprising a client-side CKKS encryption module, a cloud-side HE inference engine, and a distributed key management layer, ensuring that raw input data is never exposed to the server at any stage of computation. The framework introduces HE-compatible polynomial activation function approximations via degree-5 Chebyshev minimax polynomials, an optimized SIMD ciphertext batching strategy exploiting Ring Learning With Errors (RLWE) slot packing, and an adaptive lazy bootstrapping pipeline to substantially reduce homomorphic evaluation depth and inference latency. A formal security analysis under the IND-CPA model grounded in the RLWE hardness assumption demonstrates resistance to inference, model inversion, and membership inference attacks. Comprehensive experiments across three domains benchmark image classification (MNIST: 99.28%, CIFAR-10: 90.37%), medical imaging (93.61%), and financial fraud detection (96.44%) demonstrate that HE-CloudML achieves near-plaintext accuracy with a maximum accuracy drop of 1.81%, while delivering up to 26.9× latency improvements over CryptoNets.