Homomorphic Encryption & Federated Learning for Secure AI

The whitepaper highlights the cryptographic technology which is Homomorphic Encryption (HE) integrated with Federated Learning (FL) while training machine learning (ML) models to provide a transformative approach to one of the biggest challenges faced by organisations in the current industrial scenario which is confidentiality of sensitive information and data privacy. Implementing FL can be achieved using multiple remotely based and decentralised local server models to train machine learning models while computing only the encrypted data from the server side.

The problem revolves around:

• Ensuring data privacy when training ML models since raw data is used at large which compromises sensitive data and integrity.
• Hindrance of trust, reducing productivity, violating protocols and regulatory noncompliance.
• Accuracy, scalability, and the reliability of the data which often plagues this process.

To overcome these challenges, we propose Federated Learning with Homomorphic Encryption:

Federated Learning allows the machine learning model to use real-time data updates from the local models installed remotely while also maintaining data integrity and safeguarding sensitive information with the implementation of Homomorphic Encryption.

• HE facilitates curated utilisation of the encrypted data to perform mathematical operations without compromising the data privacy or violating protocols when analysing the data from datasets with similar features but different samples.
• This further helps in alleviating data breaches, foster trust, and attenuate risks.
• However, some of the limitations consist of higher computational overheads, bottlenecking on older GPUs.

This paper outlines the key components of the proposed Federated Learning (FL) model and highlights the advantages and effectiveness of the said model while keeping the data integrity intact and working with the cipher values to perform computations and train the ML model accordingly.