fairlearn-fhe

Drop-in encrypted Fairlearn metrics. Fairlearn-compatible API for the supported metric subset; ciphertext arithmetic over CKKS via TenSEAL.

fairlearn-fhe is an early-stage project maintained at https://github.com/BAder82t/fairlearn-fhe.

# plaintext
from fairlearn.metrics import demographic_parity_difference
disp = demographic_parity_difference(y_true, y_pred, sensitive_features=A)

# encrypted (one import change)
from fairlearn_fhe.metrics import demographic_parity_difference
from fairlearn_fhe import build_context, encrypt
ctx = build_context()
y_p_enc = encrypt(ctx, y_pred)
disp = demographic_parity_difference(y_true, y_p_enc, sensitive_features=A)

disp is numerically equivalent to the plaintext result within CKKS noise tolerance (< 1e-4 abs error in default settings).

Trust models

Two modes are supported. The default ports the regaudit-fhe convention; the second goes further.

Mode A — encrypted predictions, plaintext sensitive features (default)

• Encrypted: y_pred.
• Plaintext: y_true, sensitive_features, group counts.
• Cost: one ct×pt multiply + slot-sum per group → depth 1.

Mode B — fully-encrypted predictions and sensitive features

from fairlearn_fhe import build_context, encrypt, encrypt_sensitive_features
from fairlearn_fhe.metrics import demographic_parity_difference

ctx = build_context()
y_pred_enc = encrypt(ctx, y_pred)
sf_enc     = encrypt_sensitive_features(ctx, sensitive_features, y_true=y_true)

disp = demographic_parity_difference(y_true, y_pred_enc, sensitive_features=sf_enc)

• Encrypted: y_pred, the per-row group-membership masks.
• Plaintext (auditor metadata): group counts, per-group positive/negative counts (passed via y_true= at encryption time).
• Cost: ct×ct + ct×pt + slot-sum per group → depth 2.

y_true remains plaintext in both modes (it is the auditor's ground truth). The denominators of TPR/FPR-style metrics — per-group positive/negative counts — are always revealed: there is no fairness signal without them.

Per-group rates are decrypted at the audit boundary; final aggregation (max, min, ratio, difference) runs on those K plaintext scalars.

Supported metrics

Plus MetricFrame.fhe() returning an EncryptedMetricFrame.

Backends

Two CKKS backends share a single API:

from fairlearn_fhe import build_context

ctx_tenseal = build_context(backend="tenseal")  # default; pip-installable
ctx_openfhe = build_context(backend="openfhe")  # native OpenFHE backend, opt-in

Benchmarked on n=1024, 3 sensitive groups, depth-6 circuit:

On the included benchmark, OpenFHE gives lower numeric error; TenSEAL is faster per metric and ships via pip on every supported platform.

Install

pip install fairlearn-fhe          # tenseal backend
pip install fairlearn-fhe[openfhe] # add openfhe backend (requires C++ build)
pip install fairlearn-fhe[signing] # add Ed25519 envelope signing helpers

Verify an audit envelope without importing an FHE backend:

fairlearn-fhe-verify envelope.json

License

Apache-2.0. Compatible with Fairlearn (MIT).