In-Situ PIM Bootstrapping for Low-Latency Fully Homomorphic Encryption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fully homomorphic encryption (FHE) systems face significant performance and memory overheads due to large data size expansion and computational costs, with bootstrapping being a major bottleneck, and current CPU and GPU implementations do not provide sufficient speedup for practical applications.
Innovation Solution
Implementing a fully homomorphic encryption system using in-situ processing-in-memory (PIM) with a server PIM accelerator for operations on encrypted data, employing pipelined architectures for bootstrapping and key switching, and utilizing PIM-enabled bootstrapping to maintain low noise levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If bootstrapping is used to reduce noise growth during FHE function evaluation, then more complex operations can be performed, but latency increases by 100-1000×
Solution Approach 1:
The bootstrapping process is divided into multiple pipeline stages that can execute simultaneously. The system segments the bootstrapping operation into distinct computational steps that are distributed across multiple processing units, allowing parallel execution and reducing overall latency while maintaining the ability to perform complex FHE operations.
Solution Approach 2:
The system implements dynamic pipeline scheduling that adapts to different FHE operation requirements. The pipeline can dynamically adjust its execution flow and resource allocation based on the specific bootstrapping needs, optimizing performance for varying levels of operational complexity without requiring full pipeline reconfiguration.
2Reliability
If FHE encryption is applied to preserve data privacy, then data security is enhanced, but data size expands from 4B to more than 20 KB
Solution Approach 1:
The system merges multiple FHE ciphertext operations into single pipeline executions. By combining related FHE operations and processing them together through the unified pipeline architecture, the system reduces redundant data expansion and memory operations, thereby mitigating the data size overhead while preserving security.
Solution Approach 2:
The FHE pipeline is designed as a universal processing architecture that can handle multiple types of FHE operations (encryption, decryption, bootstrapping, key switching) through a single unified system. This multi-functionality reduces the need for separate specialized hardware for each operation, optimizing resource utilization and reducing overall system overhead.
3Productivity
If homomorphic multiplication of FHE encrypted integers is performed, then computational capability is maintained, but tens of millions of operations are required
Solution Approach 1:
The system performs preliminary preparation of FHE operands and pre-computation of intermediate values before the main multiplication operation. By pre-processing inputs and setting up computational intermediates in advance, the system reduces the number of operations required during the actual homomorphic multiplication, thereby maintaining computational capability while reducing operational complexity.
Solution Approach 2:
The FHE pipeline maintains continuous execution of useful computational actions without idle cycles. The pipeline architecture ensures that processing units remain continuously engaged in productive FHE operations, eliminating wasted computation and reducing the total number of operations needed to complete homomorphic multiplication tasks.
4Speed
If CPU and GPU implementations of RGSW-based FHE schemes are used, then processing speed is improved, but scaling is insufficient to provide adequate speedup
Solution Approach 1:
The system transitions from traditional CPU/GPU processing dimensions to a memory-centric processing dimension with PIM. By organizing FHE operations around memory architecture rather than processor architecture, the system achieves scaling in a new dimension that overcomes the limitations of conventional processors, providing both speed improvement and enhanced scalability through memory bandwidth utilization.
Data Source
AI summary
A method of operating on encrypted data can be performed by receiving ciphertexts at a server that is configured to operate on the ciphertexts generated using a 3rd generation RGSW based fully homomorphic encryption system, operating on the ciphertexts received at the server in response to requested operations to generate respective input ciphertexts including ciphertext polynomials and ciphertext integers that are representative of the input ciphertexts, and processing the input ciphertexts in a server processing-in-memory device, that is operatively coupled to the server, to perform operations on the input ciphertext using the server processing-in-memory device, in-situ.