pyfe4ai — Python Functional Encryption for AI

PyFE4AI is a research-oriented Python library of functional encryption (FE) schemes for AI security and privacy. It covers single-input (SIFE), multi-input (MIFE), multi-client (MCFE), decentralized, threshold, function-hiding, and quadratic constructions under DDH (3072-bit groups or ristretto255), Damgård-DDH, Paillier, LWE, Ring-LWE and pairing-based assumptions. Every scheme is built from its id with pyfe4ai.create("family/variant", config), which returns a key generator and a crypto class; the arguments of encryption and decryption are family-specific, and FESchemeWrapper offers a uniform layer for ML workloads. An optional Rust extension runs the hot paths natively. Check each scheme’s review status (pyfe4ai.get_scheme_status) before relying on its privacy.

Project homepage · GitHub · Issues · Changelog

Where to start

• New to the library: Installation, then Quick start.

• Picking a scheme for a task: Choosing a scheme.

• Threshold decryption and federated learning: Threshold decryption and the ML adapter.

• Coming from version 0.3.0: Upgrading from 0.3.0.

Security & status

• Review status: 42 registered schemes, 34 reviewed, 5 experimental, 3 unreviewed and 0 known-insecure. The experimental ones are the threshold layers of tMIFE-LWE, tMCFE-LWE, tMCFE-Ring-LWE and tMCFE-FH-Multi-IPE and the carry mask of dMCFE-DDH (our own constructions, no published proof). pyfe4ai.get_scheme_status("family/variant") returns a scheme’s entry.

• Security profiles: security_profile="toy", the default, is not secure; "standard" aims at 128-bit security (3072-bit DL groups and DCR moduli, lattice-estimator bounds for LWE and Ring-LWE, stat_param 128), except that pairing-based schemes get BLS12-381 (about 120-126-bit, with a FESecurityWarning) and sife/fullysec_lwe, mife/fullysec_lwe, mife/lwe_threshold and mcfe/paillier_decentralized are refused. The ABDP15 LWE schemes take matrix_mode ("memory" or "stream") to trade memory for time.

• Threshold schemes: decryption servers are honest-but-curious; a deviating server can change the result unnoticed.

• ML adapter: FESchemeWrapper runs the key generator and holds the master key, a single trust domain meant for experiments.

Details: Security & Status, and the README sections Scheme Status, Security Profiles and Known Issues / Security Notice.

Contents

• Installation
  • From source
  • The Rust extension
  • Other pairing curves
  • Running the tests
  • Building this documentation
• Quick start
  • A first example
  • Several clients
  • Federated learning
  • Real parameters
  • Finding a scheme
• Choosing a scheme
  • Families
  • Assumptions
  • Capabilities
  • Review status
  • Which scheme for which task
  • The registered schemes
• Security & Status
  • Scheme status
  • Security profiles
  • Threshold trust model
  • ML adapter
  • Gaussian sampling
  • Status API
  • Profiles API
• Configuration options
  • Server and client ids
  • Security profiles
  • Refused and constrained combinations
  • Options of every key generator
  • Options of each scheme
• Environment variables and the parameter cache
  • Environment variables
  • The parameter cache
• Threshold decryption and the ML adapter
  • Roles and flow
  • Threshold MCFE over Ring-LWE
  • Accounting of the flooding noise
  • Federated learning with the ML adapter
  • What fails, and why
• Threads, Processes and Copies
  • Selection: backend, threads, sampler
  • Objects several threads may share
  • Copies and pickles
  • Processes and fork
  • Process-wide state
• Glossary
  • Scheme taxonomy
  • Configuration keys
  • Labels, tags and keys
  • Review status and trust
  • Cryptographic variables
  • API method naming
• Upgrading from 0.3.0
  • Before you upgrade
  • Imports and names
  • Behaviour changes
  • Formats and defaults
  • Re-keying
  • Tooling
• API Reference
  • Top-level package
  • Public and internal modules

Indices and tables

• Index

• Module Index

• Search Page