BlockSecML: A Privacy-Preserving Machine Learning Framework Using Homomorphic Encryption, Advanced Encryption Standard-256 (AES-256), and Blockchain

Abstract

With machine learning being used more and more in sensitive areas like healthcare and finance, protecting the privacy of training data has become a serious concern. Most existing approaches handle encryption, secure computation, and audit logging as separate steps, which creates gaps where data can be exposed. In this paper, we present BlockSecML, a framework that brings together three technologies - Advanced Encryption Standard (AES)-256-Galois/Counter Mode encryption, Paillier partially homomorphic encryption (PHE), and Hyperledger Fabric blockchain - into a single working pipeline for privacy-preserving federated machine learning. The core design is as follows: gradient computation occurs entirely in plaintext at each data owner node; the computed gradients are immediately encrypted under Paillier PHE before any transmission; and the aggregation server performs homomorphic addition on the encrypted gradients without decrypting any individual contribution. Two brief in-memory plaintext windows exist at the data owner node - during local gradient computation and during AES-to-Paillier re-encoding - both of which are confined to the data owner's own device and do not involve network transmission. Every significant pipeline event is logged as an immutable, hash-chained record on a Hyperledger Fabric ledger. Experiments were conducted on an Intel Core i5-10210U laptop with 20 GB RAM. BlockSecML achieves classification accuracy of 97.6 ± 0.21% on MNIST (Modified National Institute of Standards and Technology), 84.9 ± 0.58% on MIMIC-III (Medical Information Mart for Intensive Care), and 91.7 ± 0.27% on a synthetic financial fraud dataset, compared to plaintext baselines of 98.9%, 87.2%, and 93.4%, respectively. Training time overhead ranges from 116% to 179% relative to plaintext training, with PHE computation accounting for approximately 87% of this overhead. All figures represent upper-bound estimates due to the single-machine experimental setup. Blockchain logging adds under 100 ms per round. The code is available at https://github.com/deeptisharda29/BlockSecML.