L4 · APPLICATIONMEASURED

Encrypted Feature Engineering

encrypted-feature-engineering · spec v0.1.0 · 0 implementations · 6 runners

Create implementationSpecification

DESCRIPTION

Applies the transformations a model expects — normalization, binning, one-hot encoding, aggregation over groups — to data that stays encrypted.

This is the step private machine learning usually skips, and skipping it is where the privacy goes: if features have to be prepared in the clear before encrypted inference, the data was exposed before the interesting part began.

Every transformation here is awkward under encryption for its own reason. Normalization needs statistics over the dataset, binning needs comparisons, one-hot encoding needs equality tests, and grouping needs data-dependent aggregation. The kernel is the pipeline; which transformations an implementation covers is its declared coverage.

INTERFACE

encrypted_feature_pipeline(
  in  ct     : Ciphertext  // the encrypted raw records
  out ct     : Ciphertext  // the encrypted prepared feature vectors
  ...               // anything else the realization needs — keys, context, encoding — is its own concern
)

PARAMETER SCHEMA

NAMETYPERANGE

recordsinteger1 – 2^24

raw_featuresinteger1 – 2^12

output_featuresinteger1 – 2^14

target_precisionnumber0.5 – 1

The schema belongs to the kernel. Each implementation declares which part of it it supports.

CORRECTNESS

01Prepared features match what the same pipeline produces in the clear, within the declared precision.

02Statistics used for scaling are computed over the declared population, not leaked from elsewhere.

03The transformations covered are stated; a pipeline that skips one is a narrower coverage, not a different result.

SECURITY PROPERTIES

Raw records never appear in the clear, including during the statistics pass. Any parameter derived from the data — a mean, a bin boundary — is itself encrypted or explicitly declared public.

ASSUMPTIONS

The schema of the raw records is public; their contents are not.

REFERENCES

repoFHERMA participation guide