Backends¶

Two CKKS backends share an identical API. Pick at context construction:

from fairlearn_fhe import build_context

ctx_tenseal = build_context(backend="tenseal")  # default
ctx_openfhe = build_context(backend="openfhe")  # opt-in

Or set the process-wide default:

from fairlearn_fhe._backends import set_default_backend
set_default_backend("openfhe")

Comparison¶

	TenSEAL	OpenFHE
Install	`pip install fairlearn-fhe`	`pip install fairlearn-fhe[openfhe]` (C++ build)
Underlying lib	Microsoft SEAL	OpenFHE 1.x
Default ring	2¹⁴	tunable, 2¹⁵ default at depth 6
Latency	faster per metric	~2× TenSEAL
Precision	~1e-7 abs err	~1e-10 abs err
Bootstrapping	no	yes (not used by fairness metrics)
Packaging	pip-installable default	native dependency, opt-in

Benchmark (n=1024, 3 sensitive groups, depth-6 circuit)¶

backend	ctx build	encrypt	dp_diff	dp abs err	eo_diff	eo abs err
tenseal	888 ms	7.5 ms	284 ms	1e-7	562 ms	2e-7
openfhe	321 ms	13.5 ms	505 ms	2e-10	1015 ms	4e-11

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

Backend-specific notes¶

TenSEAL¶

Default coeff_mod_bit_sizes: [60, 40, 40, 40, 40, 40, 40, 60] — 6 multiplicative levels at 128-bit security.
poly_modulus_degree=2**14 by default; pass 2**15 for higher precision and a longer rotation-key set.

OpenFHE¶

OpenFHE chooses the modulus chain internally given multiplicative_depth=6, scaling_mod_size=40, batch_size=N.
Pass batch_size to build_context to override the default slot count (default = poly_modulus_degree // 2).
Galois keys are pregenerated for the Halevi-Shoup ladder; EvalSum uses them for sum_all.

Dispatch internals¶

fairlearn_fhe._backends exposes a tiny registry:

from fairlearn_fhe._backends import get_backend, get_default_backend, set_default_backend

The EncryptedVector wrapper threads the active backend through every operation (add, sub, mul_pt, mul_ct, sum_all, decrypt), so every metric circuit is backend-agnostic.