Orion: A Fully Homomorphic Encryption Framework for Deep Learning

Published: March 27, 2025

Language: Английский

---

CiFlow: Dataflow Analysis and Optimization of Key Switching for Homomorphic Encryption

Published: May 5, 2024

Language: Английский

---

A User-Centered Framework for Data Privacy Protection Using Large Language Models and Attention Mechanisms

Applied Sciences, Journal Year: 2024, Volume and Issue: 14(15), P. 6824 - 6824

Published: Aug. 5, 2024

This paper introduces a user-centered data privacy protection framework utilizing large language models (LLMs) and user attention mechanisms, which are tailored to address urgent concerns in sensitive processing domains like financial computing facial recognition. The innovation lies novel mechanism that dynamically adjusts weights based on characteristics needs, enhancing the ability identify protect information effectively. Significant methodological advancements differentiate our approach from existing techniques by incorporating user-specific into traditional LLMs, ensuring both accuracy privacy. We succinctly highlight enhanced performance of this through selective presentation experimental results across various applications. Notably, computer vision, application led improved metrics over multi-head self-attention methods: FasterRCNN achieved precision, recall, rates 0.82, 0.79, 0.80, respectively. Similar enhancements were observed with SSD, YOLO, EfficientDet notable increases all metrics. In natural tasks, significantly boosted Transformer, BERT, CLIP, BLIP, BLIP2, demonstrating framework’s adaptability effectiveness. These streamlined underscore practical impact technological advancement proposed framework, confirming its superiority without compromising efficacy.

Language: Английский

---

Cinnamon: A Framework for Scale-Out Encrypted AI

Published: Feb. 3, 2025

Language: Английский

---

Secure Machine Learning Hardware: Challenges and Progress [Feature]

IEEE Circuits and Systems Magazine, Journal Year: 2025, Volume and Issue: 25(1), P. 8 - 34

Published: Jan. 1, 2025

Language: Английский

---

FAST: FPGA Acceleration of Fully Homomorphic Encryption with Efficient Bootstrapping

Published: Feb. 26, 2025

Bootstrapping is a critical operation in Fully Homomorphic Encryption (FHE) for privacy-preserving computation. Due to its significant computational overhead, accelerating bootstrapping crucial practical FHE applications involving deep evaluation circuits. In this paper, we introduce FAST, an FPGA-based accelerator efficient bootstrapping. We propose novel datapath optimizations two key operations bootstrapping: homomorphic linear transformation (HLT) and polynomial evaluation. Our memory-efficient designed HLT significantly reduces off-chip ciphertext access. also speed up the process by reducing number of required HE operations. conduct in-depth analysis Advanced Algorithm (ABA) highlight advantages. FAST first support ABA, demonstrating speedup addition, develop versatile permutation circuit handle diverse patterns FHE, achieving high throughput resource utilization. Compared with state-of-the-art (SOTA) GPU FPGA designs, achieves 8.84× 5.89× speedups bootstrapping, respectively. As illustrative examples applications, show that delivers over 20× logistic regression training compared SOTA implementation outperforms design 1.43× ResNet-20 inference.

Language: Английский

---

HEngine: A High Performance Optimization Framework on a GPU for Homomorphic Encryption

ACM Transactions on Architecture and Code Optimization, Journal Year: 2025, Volume and Issue: unknown

Published: April 28, 2025

Homomorphic encryption (HE) represents an technology that allows for direct computation on encrypted data without requiring decryption. However, the substantial computational complexity and significant latency associated with HE has impeded its broader adoption in practical applications. To address these challenges, we propose a GPU-based acceleration framework, namely HEngine, tailored homomorphic tasks. Specifically, first warp shuffle-based optimization method two key phases, i.e., inverse Chinese Remainder Theorem (ICRT) number theoretic transformation (NTT), to mitigate synchronization overhead encryption. Secondly, fuse NTT kernel inner product imbalance between memory access computation. Thirdly, considering potential difference amount of tasks users real world, design different encoding methods small batch large inference improve efficiency. Finally, experiments demonstrate our proposed framework achieves 218 × speedup multiplication compared CPU-based SEAL library. In addition, convolutional neural network shallow structures, amortized performance at millisecond level sub-millisecond data, respectively. For deeper structures (i.e., ResNet-20), second-level inference.

Language: Английский

---

LP-HENN: fully homomorphic encryption accelerator with high energy efficiency

Cybersecurity, Journal Year: 2025, Volume and Issue: 8(1)

Published: May 30, 2025

Abstract Fully homomorphic encryption (FHE) enables direct computation on encrypted data without decryption, ensuring privacy in cloud computing scenarios and preventing the leakage of sensitive information. However, computational overhead HE typically exceeds that plaintext by 4 to 5 orders magnitude, while energy consumption is 6 magnitude higher. These substantial performance overheads significantly hinder widespread adoption FHE. This paper proposed LP-HENN, a novel low-power energy-efficient FHE accelerator architecture leverages RISC-V vector coprocessor ReRAM crossbar arrays. LP-HENN targets power-constrained application such as edge devices, aiming provide highly acceleration support for applications. collaborative work processor crossbars, employing optimization strategies achieve full pipelining minimize memory access. Furthermore, this parameter selection model early-stage design, which achieves an optimal balance between through multiple parameters. Experimental results show that, FHE-based convolutional neural network (HE-CNN) inference application, 31.82Ã- 11920.56Ã- improvement efficiency, respectively, compared CPU. Compared FxHENN, state-of-the-art FPGA-based with high efficiency 2.36Ã- 10.04Ã- respectively. The comparable F1, ASIC accelerator, featuring low power design suitable computing.

Language: Английский

---

High-precision RNS-CKKS on fixed but smaller word-size architectures: theory and application

Published: Nov. 22, 2023

A prevalent issue in the residue number system (RNS) variant of Cheon-Kim-Kim-Song (CKKS) homomorphic encryption (HE) scheme is challenge efficiently achieving high precision on hardware architectures with a fixed, yet smaller, word-size bit length W , especially when scaling factor satisfies log Δ >

Language: Английский

---

Dramaton: A Near-DRAM Accelerator for Large Number Theoretic Transforms

IEEE Computer Architecture Letters, Journal Year: 2024, Volume and Issue: 23(1), P. 108 - 111

Published: Jan. 1, 2024

With the rising popularity of post-quantum cryptographic schemes, realizing practical implementations for realworld applications is still a major challenge. A bottleneck in such schemes fetching and processing large polynomials Number Theoretic Transform (NTT), which makes non Von Neumann paradigms, as near-memory processing, viable option. We, therefore, propose novel near-DRAM NTT accelerator design, called DRAMATON. Additionally, we introduce conflict-free mapping algorithm that enables DRAMATON to process NTTs with minimal hardware overhead using fixed-permutation network. achieves 5-207× speedup latency over state-of-the-art 97× improvement EDP recent accelerator.

Language: Английский

---