Skip to content

Exploring Bitslicing Architectures for Enabling FHE-Assisted Machine Learning

Authors: Soumik Sinha; Sayandeep Saha; Manaar Alam; Varun Agarwal; Ayantika Chatterjee; Anoop Mishra; Deepak Khazanchi; Debdeep Mukhopadhyay
Year: 2022
Venue: IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems

type: technique collection: ppml status: auto

Links: DOI

Summary

Auto-extracted summary (unreviewed)

The paper proposes architectural tricks for encrypted deep-learning inference built on exact (non-approximate) FHE schemes: model quantization combined with bitsliced data packing reduces the number of costly bootstrapping operations and yields a functionally correct, highly parallel inference pipeline with a moderate memory footprint. On MNIST the bitsliced pipelines run 37x faster than the non-bitsliced versions of the same architecture while needing about 700 MB of memory bandwidth.

Abstract

Homomorphic encryption (HE) is the ultimate tool for performing secure computations even in untrusted environments. Application of HE for deep learning (DL) inference is an active area of research, given the fact that DL models are often deployed in untrusted environments (e.g., third-party servers) yet inferring on private data. However, existing HE libraries [somewhat (SWHE), leveled (LHE) or fully homomorphic (FHE)] suffer from extensive computational and memory overhead. Few performance optimized high-speed homomorphic libraries are either suffering from certain approximation issues leading to decryption errors or proven to be insecure according to recent published attacks. In this article, we propose architectural tricks to achieve performance speedup for encrypted DL inference developed with exact HE schemes without any approximation or decryption error in homomorphic computations. The main idea is to apply quantization and suitable data packing in the form of bitslicing to reduce the costly noise handling operation, Bootstrapping while achieving a functionally correct and highly parallel DL pipeline with a moderate memory footprint. Experimental evaluation on the MNIST dataset shows a significant (37×) speedup over the nonbitsliced versions of the same architecture. Low memory bandwidths (700 MB) of our design pipelines further highlight their promise toward scaling over larger gamut of Edge-AI analytics use cases.
Source: pdf

Extracted details

Field Value Source Confidence
contribution bitslicing-based data packing for exact FHE deep-learning inference llm 0.60
contribution quantization to reduce bootstrapping cost without approximation or decryption errors llm 0.60
contribution highly parallel encrypted DL pipeline with moderate memory footprint llm 0.60
schemes TFHE llm 0.60
libraries TFHE llm 0.60
datasets MNIST llm 0.60
tasks encrypted deep-learning inference llm 0.60
tasks image classification llm 0.60
tasks Edge-AI analytics llm 0.60
threat model untrusted third-party cloud server performing inference on private data llm 0.60
interaction non-interactive FHE llm 0.60
code url https://github.com/qianlou/SHE (URL found in the full text) regex 0.50
code url https://github.com/tfhe/tfhe (URL found in the full text) regex 0.50
code url https://github.com/zama-ai/concrete (URL found in the full text) regex 0.50
llm relevance medium llm 0.60
key result speed-up: 37x (MNIST, bitsliced vs non-bitsliced version of the same architecture) llm 0.60
key result memory bandwidth: 700 MB (design pipelines) llm 0.60

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: Leveled / somewhat HE 17 [p.1], TFHE 9 [p.3], CKKS 2 [p.1], Paillier 2 [p.12], MPC 2 [p.1]
  • library: Concrete (Zama) 12 [p.2], Zama 2 [p.3], Microsoft SEAL 1 [p.12]
  • model: CNN 26 [p.2], AlexNet 1 [p.12]
  • dataset: MNIST 15 [p.1], CIFAR-10 1 [p.10]
  • hardware: GPU 5 [p.10]
  • technique: Bitslicing 57 [p.1], Bootstrapping 34 [p.1], Quantization 32 [p.1], Convolution 20 [p.2], ReLU 6 [p.4], Softmax 2 [p.5], Sigmoid 2 [p.4], Square activation 2 [p.2], SIMD packing 2 [p.5]
  • assumption: LWE 8 [p.4], RLWE 4 [p.5]
  • task: IoT / edge 3 [p.10]
  • topic: Homomorphic encryption 125 [p.1], Private / secure inference 6 [p.2], Open source 5 [p.10]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/ppml/inference/sinha-2022_exploring-bitslicing-architectures-for-fhe-assisted-ml.pdf ppml/inference 12 3.6 MB d65dfccc8807 was Sinha et al. - 2022 - Exploring Bitslicing Architectures for Enabling FHE-Assisted Machine Learning.pdf

Database key: exploring-bitslicing-architectures-for-enabling-fhe-assisted-machine · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set exploring-bitslicing-architectures-for-enabling-fhe-assisted-machine COLUMN VALUE