TY  - EJOU
AU  - Ahmed, Saadaldeen Rashid 
AU  - Siryeh, Fatima Abu 
AU  - Yadkar, Mohammed Shamar 
AU  - Bayat, Oguz 
AU  - Miah, Abu Saleh Musa 
AU  - Farid, Fahmid Al 
AU  - Karim, Hezerul Abdul 

TI  - GuardML: A Hybrid Homomorphic Encryption Framework for Privacy-Preserving Federated Deep Learning
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Privacy-preserving federated learning (FL) has emerged as an effective paradigm for collaborative model training across distributed data sources while maintaining data confidentiality. However, protecting sensitive information during model aggregation remains a significant challenge in distributed deep learning environments. This paper introduces GuardML, a secure federated learning framework that integrates a Recurrent Neural Network (RNN) with Hybrid Homomorphic Encryption (HHE) to enable privacy-preserving learning on high-dimensional distributed datasets. In the proposed system, client nodes perform local model training using encrypted data representations, ensuring that raw data remains protected during the learning process. Encrypted model updates are transmitted to a central server, where secure aggregation is performed without exposing individual client information. Experimental evaluation demonstrates that GuardML achieves a classification accuracy of 98.79%, outperforming several federated learning baselines, including KNN (89.27%), SVM (93.38%), and K-Means (82.86%), combined with Differential Privacy (DP) or Secure Multi-Party Computation (SMPC) mechanisms. Although HHE introduces additional encryption overhead (approximately 120 ms) compared with DP-based approaches (60–80 ms), the framework maintains low inference latency (65 ms) and efficient encrypted communication (12.4 MB). The proposed client–server architecture maintains encryption throughout the learning pipeline, mitigating risks such as adversarial inference, model inversion, and membership inference attacks. The study further analyzes the trade-offs among encryption complexity, computational efficiency, and communication cost in secure federated learning systems. The results demonstrate that hybrid homomorphic encryption can provide strong privacy guarantees while maintaining high predictive performance for distributed machine learning tasks. Future work will focus on improving computational efficiency through GPU-accelerated homomorphic encryption, hybrid cryptographic optimization, and model compression techniques to support scalable privacy-preserving AI applications in distributed data environments.
KW  - Federated learning; hybrid homomorphic encryption; privacy-preserving machine learning; distributed deep learning; secure model aggregation; recurrent neural networks (RNN)

DO  - 10.32604/cmc.2026.082782