How google/fully-homomorphic-encryption works
This tree is a collection of FHE demonstration workloads, not the HEIR compiler itself. The intended user journey is to prepare or export a model, compare cleartext inference with encrypted evaluation, and inspect the resulting inference or performance. The README describes model conversion, multiple FHE schemes/backends, and encrypted computation, but only go.mod was sampled; the demo-specific connections below are therefore dashed and unverified.
Main entry points: evaluate_fhe.go Encrypted evaluation · export_mlir.py MLIR export · evaluate_cleartext.py Cleartext evaluation · train.py Model training
Components 16
Demo Workloads
- Fraud detection
demos/cc_fraud - Criteo inference
demos/criteo - Hotword inference
demos/hotword - MNIST inference
demos/mnist - Network anomaly
demos/network_anomaly
Model Preparation
- Model training
demos/cc_fraud/train/train.py - Data preparation
demos/cc_fraud/utils - MLIR export
demos/cc_fraud/torch/export_mlir.py
Evaluation
- Cleartext evaluation
demos/cc_fraud/cleartext/evaluate_cleartext.py - Encrypted evaluation
demos/cc_fraud/lattigo/evaluate_fhe.go - Evaluation suites
demos/cc_fraud/lattigo/evaluate_fhe_suite.go - Timing analysis
demos/cc_fraud/lattigo/evaluate_fhe_timing.go - Decrypted result
FHE Backends
- Lattigo backend
demos/cc_fraud/lattigo - OpenFHE backend
demos/cc_fraud/openfhe
Other components
- Demo user
Connections 14
- Demo user → Fraud detection: runs demo
- Demo user → Criteo inference: runs demo
- Demo user → Hotword inference: runs demo
- Demo user → MNIST inference: runs demo
- Demo user → Network anomaly: runs demo
- Model training → MLIR export: exports model
- Data preparation → Cleartext evaluation: supplies data
- MLIR export → Encrypted evaluation: supplies model
- Cleartext evaluation → Encrypted evaluation: compares inference
- Encrypted evaluation → Lattigo backend: uses backend
- Encrypted evaluation → OpenFHE backend: uses backend
- Encrypted evaluation → Decrypted result: decrypts output
- Demo user → Evaluation suites: runs suite
- Demo user → Timing analysis: measures runtime