vkatg/streaming-phi-deidentification-benchmark

Streaming PHI De-Identification Benchmark

Most PHI de-identification benchmarks evaluate a single document in isolation. That is not how clinical data actually moves. A patient's name appears in a clinical note, then in an ASR transcript ten minutes later, then in imaging metadata an hour after that. Each event looks low-risk on its own. The cumulative exposure across modalities is what creates re-identification risk.

This dataset captures that. Every record is fully synthetic. It models how re-identification risk builds across a longitudinal multimodal stream and evaluates whether a masking policy responds to that accumulation or ignores it.

The short answer from the results: static policies cannot do both things at once. A policy that redacts everything achieves perfect privacy but destroys utility. A policy that pseudonymizes everything preserves utility but fails at high risk. The adaptive controller is the only one that clears both bars simultaneously.

Quick Start

import json
import pandas as pd

with open("audit_log.jsonl") as f:
    events = [json.loads(line) for line in f]

with open("audit_log_signed_adaptive.jsonl") as f:
    adaptive = [json.loads(line) for line in f]

df = pd.read_csv("policy_metrics.csv")
print(df)

Via the Hugging Face datasets library:

from datasets import load_dataset

ds     = load_dataset("vkatg/streaming-phi-deidentification-benchmark")
signed = load_dataset("vkatg/streaming-phi-deidentification-benchmark", "signed")
cm     = load_dataset("vkatg/streaming-phi-deidentification-benchmark", "crossmodal")

Why This Dataset Exists

Every open PHI benchmark we found had the same structure: a document comes in, you mask it, you score it, done. That works for NER-style de-identification. It does not work for evaluating a system where the threat model is cumulative exposure.

The specific gaps this dataset fills:

No memory across events. i2b2 and PhysioNet evaluate each record independently. A system that correctly masks event 1 and event 17 in isolation can still leak identity when you join them. This dataset tracks exposure state across the full stream and scores masking decisions against that accumulated state.

Single modality. Existing benchmarks are clinical text only. Real healthcare pipelines include ASR transcripts, imaging metadata, physiological waveforms, and audio. Cross-modal PHI linkage is a distinct threat that text-only evaluation misses entirely.

No open access. i2b2 and PhysioNet require data use agreements, which is appropriate given they contain real patient data. But it creates friction for anyone building or evaluating a de-identification system who just needs a benchmark they can run locally. This dataset is MIT-licensed, fully synthetic, and requires no agreement to use.

No adversarial coverage. None of the standard benchmarks include an attacker model. This dataset includes a formal sub-threshold probing scenario where an attacker spaces PHI submissions to stay below individual risk thresholds while accumulating cross-modal links. The adaptive controller detects and escalates on these; static policies do not.

Risk Timeline

Risk accumulates as PHI exposure is recorded across events. Patient A (research consent, ceiling: pseudo) and Patient B (standard consent, ceiling: redact) alternate. Dots are colored by the policy applied at each event. The controller escalates as risk crosses 0.40, 0.60, and 0.80.

Re-identification Risk Reduction

Delta-AUROC measures the reduction in a logistic-regression re-identifier's ability to distinguish patients from masked vs. original text, computed on a rolling 32-event window. Negative values mean the masking is working. The multi-run mean is -0.9167 ± 0.0000 (95% CI, n=10). First non-zero signal appears at event 7, once the buffer has enough samples with 2 distinct patient labels.

Privacy vs Utility — Bursty Workload

The bursty workload is the hardest case: the system has to protect high-risk returning patients while preserving utility for newly admitted low-risk ones at the same time. The target region (top-right, shaded) requires Privacy@HighRisk above 0.85 and Utility@LowRisk above 0.50. No static policy reaches it. Adaptive does.

Full Pareto frontier across all three workloads is in EXPERIMENT_REPORT.md.

Baseline Policy Comparison

Three workloads: monotonic (risk accumulates continuously), bursty (new low-risk patients enter every 6 events), mixed (70% routine / 30% high-complexity). Full numbers are in baseline_comparison.csv.

Bursty workload (primary claim)

Policy	Privacy@HighRisk	Utility@LowRisk	Consent Viols	Latency (ms)
Always-Raw	0.0000	1.0000	0	0.5
Always-Weak	0.0035	0.8466	0	1.0
Always-Synthetic	0.5642	0.6755	0	2.0
Always-Pseudo	0.8547	0.4399	0	1.5
Always-Redact	1.0000	0.0000	17	1.0
Adaptive	0.9907	0.8466	10	1.1

Monotonic workload

Policy	Privacy@HighRisk	Utility@LowRisk	Consent Viols
Always-Pseudo	0.8551	0.0000	0
Always-Redact	1.0000	0.0000	17
Adaptive	0.9863	0.0000	14

Mixed workload

Policy	Privacy@HighRisk	Utility@LowRisk	Consent Viols
Always-Pseudo	0.8544	0.4577	0
Always-Redact	1.0000	0.0000	17
Adaptive	0.9848	0.8526	6

Threshold Sensitivity

remask_thresh controls when the controller triggers pseudonym re-tokenization. Lower values trigger remask more aggressively. The chart shows how Privacy@HighRisk and Utility@LowRisk shift on the bursty workload across a grid of threshold values. The default (0.68, marked) sits at an inflection point: privacy is near ceiling and utility has not yet dropped. Full data for all three workloads is in data/threshold_sensitivity.csv.

Comparison with Existing PHI Datasets

Dataset	Modality	Access	Real Data	Streaming	Adversarial	Size
i2b2 2014	Text only	DUA required	Yes	No	No	1,304 records
PhysioNet deid	Text only	DUA required	Yes	No	No	2,434 notes
This dataset	Text, ASR, Image, Waveform, Audio	Open (MIT)	No	Yes	Yes	1K-10K events

Dataset Structure

Three configs:

Config	File	Description
default	audit_log.jsonl	Full run log, adaptive policy
signed	audit_log_signed_adaptive.jsonl	Cryptographically signed adaptive-only events
crossmodal	data/crossmodal_train.jsonl	Cross-modal link benchmark scenarios (A-E)

Fields

Field	Description
event_id	Event identifier within the stream
patient_key	Synthetic subject identifier
modality	text, asr, image_proxy, waveform_proxy, audio_proxy
chosen_policy	Policy actually applied
reason	Selection reason from the controller
risk	Cumulative re-identification risk at decision time
localized_remask_trigger	Whether risk crossing remask_thresh triggered pseudonym versioning
latency_ms	Decision latency in milliseconds
leaks_after	Residual PHI leakage post-masking
policy_version	Policy version token
consent_status	ok or capped
extra.delta_auroc	Rolling delta-AUROC at this event
extra.crdt_risk	CRDT-backed graph risk at this event
decision_blob	Full decision record: risk components, DCPG state, cross-modal matches

Sample Record

{
  "event_id": "evt_8",
  "patient_key": "A",
  "modality": "text",
  "chosen_policy": "pseudo",
  "reason": "medium-high risk: pseudonymization",
  "risk": 0.8783,
  "localized_remask_trigger": false,
  "latency_ms": 17.01,
  "leaks_after": 0,
  "consent_status": "capped",
  "extra": {
    "delta_auroc": -0.125,
    "crdt_risk": 0.512
  },
  "decision_blob": {
    "rl_policy": "redact",
    "rl_source": "rl_model",
    "rl_reward": 0.63
  }
}

Cross-Modal Benchmark Scenarios

Scenario	What it tests
A	Baseline: single-modality accumulation, no cross-modal link
B	Cross-modal link fires mid-sequence (text then ASR)
C	Three-modal convergence (text, ASR, image_proxy)
D	Remask trigger: risk crosses 0.68 mid-run, pseudonym versioning fires
E	Adversarial sub-threshold probing with cross-modal spike every 5th event

Each scenario runs all five policy variants. 260 total rows.

Risk Model

R = 0.8 * (1 - exp(-k * units)) + 0.2 * recency + link_bonus

k_units = 0.05. Link bonus: +0.20 for 2 or more cross-modal links, +0.30 for 3 or more. Validated against a closed-form combinatorial reconstruction probability (Pearson r = 0.881, n=34). Cross-modal PHI correlation between image and audio modalities is r = 0.081, confirming largely independent signals and validating the CROSS_MODAL_SIM_THRESHOLD = 0.30 design constant.

Consent Layer

All consent-cap events occur on patient A (research consent, ceiling: pseudo): the controller decided redact at high risk, downgraded to pseudo by the consent layer. Patient B (standard consent, ceiling: redact) was never capped. 14 caps total across 34 live events. Full event-by-event log is in consent_cap_log.jsonl.

RL Agent

PPO with an LSTM-backed policy network (128-dim hidden, 2 layers, 14-dimensional state), pretrained for 200 stratified episodes. Overall reward mean: 0.2806 (min -0.3337, max 0.7093, n=234). Warmup mean: 0.4994 (n=22). Model-driven mean: 0.2579 (n=212). All 34 live-loop events were model-driven. Reward statistics are in rl_reward_stats.json.

CRDT Federation

Federated graph merge demo: two edge devices, device_A (text + image_proxy) and device_B (audio_proxy + text). Post-merge: 3 nodes, merged risk = 0.1806, convergence guaranteed. Full output in dcpg_crdt_demo.json.

System Architecture

Module	Description
context_state.py	Per-subject PHI exposure state, persisted via SQLite
controller.py	Risk scoring and adaptive policy selection
dcpg.py	Dynamic Contextual Privacy Graph: cross-modal identity linkage
dcpg_crdt.py	CRDT-based graph merging for distributed/federated deployments
dcpg_federation.py	Streaming delta sync and HMAC-authenticated inter-device federation
cmo_registry.py	Composable Masking Operator registry and DAG execution
cmo_media.py	Token-level synthetic replacement ops (names, dates, MRNs, facilities)
flow_controller.py	DAG-based policy flow controller with audit provenance
masking.py	High-level masking dispatcher: routes events to the right CMO chain
masking_ops.py	Low-level masking primitives per policy tier
rl_agent.py	PPO reinforcement learning agent for adaptive policy control
phi_detector.py	PHI span detection and leakage measurement
consent.py	Consent token resolution and policy ceiling enforcement
downstream_feedback.py	Rolling utility monitor for downstream task signal
metrics.py	Leakage scoring, utility proxy, and delta-AUROC computation
eval.py	Evaluation harness: latency summarization and per-policy scoring
audit_signing.py	ECDSA signing, Merkle chain, and FHIR export of audit records
baseline_experiment.py	Static policy baselines, Pareto frontier, and workload sweep
db.py	SQLite connection management and WAL configuration
schemas.py	Shared dataclasses: PHISpan, DataEvent, DecisionRecord, AuditRecord
run_demo.py	End-to-end demo: pretraining, live loop, evaluation, report generation

Files

File	Description
audit_log.jsonl	Full run log, adaptive policy
audit_log_signed_adaptive.jsonl	Signed adaptive-only audit trail
data/crossmodal_train.jsonl	Cross-modal scenario benchmark (260 rows)
data/leakage_breakdown.jsonl	Per-event leakage by PHI entity type (mrn, date, name, facility)
data/risk_trace.jsonl	Per-event risk component history (units_factor, recency, link_bonus)
data/threshold_sensitivity.csv	Privacy/utility across remask_thresh grid (0.50-0.80), all workloads
policy_metrics.csv	Per-policy metrics from the live run
latency_summary.csv	Latency distribution summary
baseline_comparison.csv	Full baseline results across all workloads
statistical_robustness.json	Multi-run (n=10) robustness statistics
controller_config.json	Controller configuration for this run
rl_reward_stats.json	PPO reward statistics
consent_cap_log.jsonl	Per-event consent cap decisions
delta_auroc_log.jsonl	Per-event delta-AUROC and CRDT risk
dcpg_crdt_demo.json	CRDT federation merge output
EXPERIMENT_REPORT.md	Full experiment report

Limitations

This dataset is fully synthetic. It models the structural properties of longitudinal healthcare streams but does not contain real clinical language, real diagnoses, or real patient records. Risk scores, leakage values, and policy decisions are generated by the simulation, not extracted from real deployments. Any system trained or evaluated here should be validated against real clinical data before production use. English-language proxies only.

Reproduce

git clone https://github.com/azithteja91/phi-exposure-guard.git
cd phi-exposure-guard
pip install -e .
python -m amphi_rl_dpgraph.run_demo

Generate the supplementary data files:

python scripts/generate_crossmodal_train.py
python scripts/generate_leakage_breakdown.py
python scripts/generate_risk_trace.py
python scripts/generate_threshold_sensitivity.py

GitHub: azithteja91/phi-exposure-guard

HF Space: vkatg/amphi-rl-dpgraph

Colab:

Citation

Cite via the CITATION.cff file in the GitHub repository.

License

MIT