Sign in Subscribe
Artificial neural networks

Advances in Homomorphic Encryption for Secure Machine Learning Inference

Strategic Insights

— 1 min read

Researchers at Isik University have published a groundbreaking study on homomorphic encryption (HE) for secure machine learning inference in sensitive environments such as healthcare and finance. Their research focuses on efficiently handling non-linear activation functions in artificial neural networks (ANNs) under homomorphic encryption. The study introduces a lightweight, ANN-based estimator that accurately approximates activation functions, achieving superior accuracy with lower computational overhead.

Key Takeaways:

  • The researchers developed a lightweight, ANN-based estimator to accurately approximate non-linear activation functions (Sigmoid and Tanh) under homomorphic encryption, significantly outperforming conventional methods.
  • The proposed estimator achieved notable improvements in accuracy, with a 2% increase for Sigmoid and up to 73% for Tanh functions, and a 96% reduction in Mean Square Error (MSE) compared to polynomial approximations.
  • The estimator was trained on plaintext data and seamlessly integrated into encrypted inference pipelines, enhancing F1-scores by approximately 2% for Sigmoid and up to 88% for Tanh, and achieving an accuracy of 97.70% and an AUC of 0.9997 on the MNIST dataset.
  • The study demonstrated the practical effectiveness and computational feasibility of the ANN estimator, making it suitable for secure and efficient ANN inference in encrypted environments.
  • The research was conducted by Mhd Raja Abou Harb, Computer Engineering Department, Isik University, Istanbul, Turkey, in collaboration with Baris Celiktas.

Statistics:

  • The proposed ANN estimator achieved an accuracy of 97.70% and an AUC of 0.9997 on the MNIST dataset.
  • The estimator enhanced F1-scores by approximately 2% for Sigmoid and up to 88% for Tanh functions.
  • The Mean Square Error (MSE) was reduced by up to 96% compared to polynomial approximations.
  • The study demonstrated significant improvements in accuracy (up to 73% for Tanh functions) and F1-scores (up to 88% for Tanh functions).

Sources:

  • ANN Activation Function Estimators for Homomorphic Encrypted Inference. IEEE Access, 2025,13():103512-103530. (IEEE Access - http://ieeexplore.ieee.org/servlet/opac?punumber=6287639).
  • Isik University.

Note: The sources provided are verbatim from the original text, with no modifications or additions.

Read more