hifd

HiFD is a holistic metric for evaluating face de-identification methods. It combines privacy (face-recognition resistance), perceptual quality, and three levels of utility preservation (macro cues, fine-grained cues, physiological signals) into a single composite score.

This package provides the scoring core: given JSON files of pre-computed estimator predictions on (a) original images and (b) de-identified outputs, hifd produces per-sample agreement scores, per-method utility levels, ensemble privacy, quality, and composite HiFD scores under three application profiles (Privacy-First, Balanced, Clinical).

The package does not run the estimators themselves — you bring your own predictions in the documented JSON schema.

Install

pip install hifd

Runtime requirements: Python ≥ 3.10, numpy, pandas, pyyaml.

Quickstart (CLI)

hifd run-all \
    --methods    methods.yaml \
    --estimators estimators.yaml \
    --data-dir   predictions/ \
    --out        scores/

Produces:

scores/
├── per_sample/per_sample_agreements.csv
├── per_method/per_method_aggregated.csv
├── per_method/method_level.csv
├── tables/composite.csv
└── tables/composite.tex

Quickstart (Python API)

import hifd
import numpy as np

# Per-task primitives (scalar)
hifd.score_age(35.0, 36.0)                 # 0.99
hifd.score_gaze([0, 0, 1], [0, 0.1, 1])    # ≈ 0.937

# Batch variants for scalar-input primitives
hifd.batch_score_age(np.array([35.0]), np.array([36.0]))  # ndarray([0.99])
# Also: batch_score_categorical, batch_score_hr, batch_privacy_score,
# batch_quality_score, batch_U1, batch_U2, batch_U3
# (Geometric primitives — score_landmark, score_gaze, score_bvp — have no
# batch form because their inputs are per-sample variable-shape arrays.)

# Composite under a profile
hifd.composite(P=0.7, Q=0.8, U1=0.6, U2=0.5, U3=0.4,
               weights=hifd.DEFAULT_PROFILES["balanced"])

# Full pipeline
r = hifd.run_pipeline(
    methods_config="methods.yaml",
    estimators_config="estimators.yaml",
    data_dir="predictions/",
    out_dir="scores/",
)
print(r.composite_df)

Input format

See docs/schema.md for the JSON envelope spec and one example per task type.

Defaults

Constants and profiles ship as immutable mappings:

hifd.DEFAULT_CONSTANTS["tau_HR"]            # 20.0
hifd.DEFAULT_PROFILES["balanced"]           # {"P": 0.2, "Q": 0.2, "U1": 0.2, ...}

To override, build a plain dict and pass it (run_pipeline(constants=...)) or provide --constants my.yaml on the CLI.

Citation

If you use HiFD in published work, please cite the paper [TBD].