Skip to content

Building Privacy-Preserving AI system: AI Inference on Encrypted Data with Fully Homomorphic Encryption

Authors: Riccardo Magni; Carlo Tassi; Emanuele D’Agostini
Year: 2025
Venue: 2025 International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA)

type: application collection: ppml status: auto

Links: DOI

Summary

Auto-extracted summary (unreviewed)

An industry (Almaviva) study that applies Zama's Concrete ML, a TFHE-based machine learning framework, to real-world-style scenarios and evaluates reliability, practical performance and production feasibility of inference on encrypted data compared with inference on unencrypted models. The authors conclude that FHE is now mature enough for some real-world use, although computation time, circuit construction, key management and encrypted data transfer impose significant costs.

Abstract

Artificial intelligence (AI) has revolutionized busi- ness strategies by enhancing service personalization and decision- making processes. However, deploying these technologies requires significant computational power, often sourced from third-party cloud services. These services facilitate complex calculations for training and inferring machine learning models without requiring substantial hardware investments. The risk of data exposure or unauthorized access remains high, particularly for sensitive and confidential information. Fully Homomorphic Encryption (FHE) [1] offers a promising solution by enabling mathematical operations on encrypted data without decryption, ensuring total data confidentiality and regulatory compliance. This paper aims to explore the application of an FHE-based Machine Learning framework to real-world scenarios, focusing on reliability, practical performance and feasibility of use in production environments, in comparison with building and mak- ing inferences on unencrypted models. The framework selected for this study is Zama Concrete ML [2], a machine learning framework based on Fully Homomorphic Encryption over the Torus (TFHE) [3]. The results obtained demonstrate the progress achieved by FHE, indicating that this technology is now mature enough to be applied in real-world contexts. Although the high costs associated with this technology — including computation time, circuit construction, key management, and encrypted data transfer — can be significant, in some cases it enables computations that would otherwise be impossible due to the inability to share sensitive data with third-party service providers.
Source: pdf

Extracted details

Field Value Source Confidence
contribution practical evaluation of Concrete ML (TFHE) for AI inference on encrypted data llm 0.60
contribution comparison of encrypted versus plaintext model building and inference in production-like settings llm 0.60
contribution assessment of reliability, performance and feasibility of FHE-based ML in real-world contexts llm 0.60
contribution discussion of deployment costs: computation time, circuit construction, key management and data transfer llm 0.60
schemes TFHE llm 0.60
libraries Concrete ML llm 0.60
libraries Concrete llm 0.60
libraries Hummingbird llm 0.60
tasks machine learning inference on encrypted data llm 0.60
tasks privacy-preserving MLaaS on third-party cloud llm 0.60
security level 128-bit llm 0.60
threat model untrusted third-party cloud service provider llm 0.60
interaction non-interactive FHE llm 0.60
code url https://github.com/microsoft/hummingbird (URL found in the full text) regex 0.50
limitations high computation time for encrypted inference llm 0.60
limitations cost of circuit construction/compilation llm 0.60
limitations key management overhead (FHE key sets for 128-bit security can be very large) llm 0.60
limitations encrypted data transfer overhead llm 0.60
llm relevance low llm 0.60

Keyword mentions

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

  • scheme: TFHE 5 [p.1]
  • library: Concrete (Zama) 14 [p.1], Concrete ML 11 [p.1], Zama 11 [p.1], ONNX 7 [p.3], PyTorch 2 [p.2]
  • model: Gradient boosting / XGBoost 4 [p.3], Random forest 2 [p.3]
  • hardware: AMD EPYC 1 [p.3], ARM 1 [p.3]
  • technique: Quantization 17 [p.2], Bootstrapping 4 [p.1], Compiler 4 [p.2], Rotations 2 [p.5]
  • assumption: Lattice-based 2 [p.1]
  • topic: Homomorphic encryption 31 [p.1], Open source 2 [p.2]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/ppml/inference/magni-2025_building-privacy-preserving-ai-system-inference-on-encrypted-data-with-fhe.pdf ppml/inference 6 469.0 KB 39f131a01e10 was Magni et al. - 2025 - Building Privacy-Preserving AI system AI Inference on Encrypted Data with Fully Homomorphic Encrypt.pdf

Database key: building-privacy-preserving-ai-system-ai-inference-on-encrypted-data · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set building-privacy-preserving-ai-system-ai-inference-on-encrypted-data COLUMN VALUE