International business machines corporation (20240256850). NEURAL NETWORK INFERENCE UNDER HOMOMORPHIC ENCRYPTION simplified abstract

NEURAL NETWORK INFERENCE UNDER HOMOMORPHIC ENCRYPTION

Organization Name

international business machines corporation

Inventor(s)

Omri Soceanu of Haifa (IL)

Nir Drucker of Zichron Yaakov (IL)

Subhankar Pal of White Plains NY (US)

Roman Vaculin of Larchmont NY (US)

Kanthi Sarpatwar of Briarcliff Manor NY (US)

Alper Buyuktosunoglu of White Plains NY (US)

Pradip Bose of Yorktown Heights NY (US)

Hayim Shaul of Kfar Saba (IL)

Ehud Aharoni of Kfar Saba (IL)

James Thomas Rayfield of Ridgefield CT (US)

NEURAL NETWORK INFERENCE UNDER HOMOMORPHIC ENCRYPTION - A simplified explanation of the abstract

This abstract first appeared for US patent application 20240256850 titled 'NEURAL NETWORK INFERENCE UNDER HOMOMORPHIC ENCRYPTION

Simplified Explanation

A trained neural network is split into a client-side portion and a server-side portion, with the client-side portion handling the initial layers and the server-side portion handling the later layers. The network is trained using specific data, and an encrypted output is computed by the server-side portion based on an encrypted input from the client-side portion.

• Trained neural network split into client-side and server-side portions
• Client-side portion handles initial layers, server-side portion handles later layers
• Network trained using specific data
• Encrypted output computed by server-side portion based on encrypted input from client-side portion

Key Features and Innovation

- Partitioning a trained neural network into client-side and server-side portions - Homomorphically encrypted intermediate result input and output - Improved security and privacy in neural network computations

Potential Applications

- Secure data processing in sensitive industries like healthcare and finance - Encrypted communication in cloud-based neural network services - Enhanced privacy protection in machine learning applications

Problems Solved

- Addressing security concerns in neural network computations - Ensuring privacy of sensitive data during processing - Preventing unauthorized access to trained neural network models

Benefits

- Enhanced data security and privacy - Secure transmission of encrypted neural network outputs - Protection against data breaches and unauthorized access

Commercial Applications

Title: Secure Neural Network Computation for Sensitive Industries This technology can be utilized in industries like healthcare, finance, and cybersecurity for secure data processing and encrypted communication. It can also be integrated into cloud-based machine learning services to enhance privacy protection and data security.

Prior Art

Further research can be conducted on homomorphic encryption techniques in neural network computations to explore existing solutions and advancements in the field.

Frequently Updated Research

Stay updated on the latest developments in homomorphic encryption for neural networks to understand the evolving landscape of secure data processing and privacy protection in machine learning applications.

Questions about Secure Neural Network Computation

How does homomorphic encryption enhance data security in neural network computations?

Homomorphic encryption allows for secure computation on encrypted data without revealing the underlying information, ensuring privacy and confidentiality in neural network operations.

What are the potential challenges in implementing homomorphic encryption in neural networks?

Implementing homomorphic encryption in neural networks may introduce computational overhead and complexity, requiring efficient algorithms and infrastructure for practical deployment.

Original Abstract Submitted

a trained neural network is partitioned into a client-side portion and a server-side portion, the client-side portion comprising a first set of layers of the trained neural network, the server-side portion comprising a second set of layers of the trained neural network, the trained neural network trained using a first set of training data. from a homomorphically encrypted intermediate result input to the server-side portion, a homomorphically encrypted output of the trained neural network is computed, the homomorphically encrypted intermediate result comprising a homomorphically encrypted output computed by the client-side portion.