loading page

An Efficient Approach for Securing Audio Data in AI Training with Fully Homomorphic Encryption
  • +1
  • Linh Nguyen,
  • Bao Phan,
  • Lan Zhang,
  • Tuy Nguyen
Linh Nguyen
Bao Phan
Lan Zhang
Tuy Nguyen

Corresponding Author:[email protected]

Author Profile


Supercomputers poised to crack current encryption standards within a decade, traditional methods face an unprecedented threat. Training artificial intelligence (AI) models on plaintext data has sparked increasing concerns regarding privacy and security. The potential risks include the possibility of data leakage or theft. To address these critical challenges, we propose a groundbreaking architecture that seamlessly integrates homomorphic encryption (HE) and AI, enabling privacypreserving analysis of sensitive audio data and paving the way for a new era of secure AI applications in audio processing. Our approach introduces novel approximations for the sigmoid (Asigmoid) and rectified linear unit (A-ReLU) activation functions, designed to be compatible with the input data distribution and to minimize the percentage of squared error (PSE) between the approximation and the original activation function, optimizing processing efficiency on encrypted data while overcoming the limitations of existing methods. We underscore the crucial role of activation function selection and HE's parameter tuning in achieving a balance between computational efficiency and model accuracy within this framework. The evaluation results using the audioMNIST and musical instrument datasets demonstrate the system's robustness with a negligible 0.04% difference between plaintext and ciphertext conditions, highlighting its promising potential for secure audio data processing in various applications.
27 Feb 2024Submitted to TechRxiv
04 Mar 2024Published in TechRxiv