Skip to content

privacy.md

Latest commit

 

History

History
89 lines (59 loc) · 3.07 KB

privacy.md

File metadata and controls

89 lines (59 loc) · 3.07 KB
Raw

Differential Privacy in LedgerLens

LedgerLens applies differential privacy (DP) noise to aggregate statistics computed during training data preparation. This prevents individual wallet activity from being reconstructed from logged or shared statistics.

Overview

privacy/dp_aggregator.py exposes a DPAggregator class that wraps standard aggregations (mean, count, histogram) with calibrated DP noise.

Two mechanisms are used:

Aggregation Mechanism Privacy guarantee
Mean Laplace Pure DP (ε, 0)
Count Laplace Pure DP (ε, 0)
Histogram Gaussian Approximate DP (ε, δ)

Sensitivity Analysis

Sensitivity quantifies the maximum change in an aggregate from adding or removing one wallet's data.

  • Mean (L1 sensitivity): Δ_mean = (feature_max − feature_min) / n
    Computed analytically per feature using known range bounds.
  • Count (L1 sensitivity): Δ_count = 1
    A single wallet can change the count by exactly 1.
  • Histogram bin count (L2 sensitivity): Δ_bin = 1
    Adding/removing one record shifts exactly one bin by ±1.

Noise Calibration

Laplace mechanism (mean, count)

Noise Lap(0, Δ/ε) is added, where:

  • Δ = sensitivity (computed above)
  • ε = DP_AGGREGATOR_EPSILON (default 1.0)

Gaussian mechanism (histogram)

Noise N(0, σ²) per bin, with:

σ = Δ · √(2 ln(1.25/δ)) / ε

where δ = DP_AGGREGATOR_DELTA (default 1e-5).

Privacy Budget Accounting

Each call to private_mean, private_count, or private_histogram consumes one unit of the configured ε (and δ for histograms). The cumulative budget is tracked in DPAggregator.budget_consumed() and logged to the training run metadata at models/dp_training_stats.json:

{
  "epsilon_used": 12.3,
  "delta_used": 3e-5,
  "queries": 37,
  "feature_dp_stats": { ... }
}

The training script logs epsilon_used, delta_used, and queries at INFO level after every run.

Epsilon Selection Guidance

Use case Recommended ε Rationale
Internal analytics (not shared) 2.0–4.0 Loose bound; utility prioritised
Shared aggregate Benford statistics 0.5–1.0 Standard DP publication threshold
High-sensitivity regulatory release 0.1–0.3 Conservative; significant noise

For sharing aggregate Benford statistics externally, ε = 1.0 is a reasonable starting point — it is a widely accepted default in the DP literature and provides meaningful noise against reconstruction attacks while preserving the directional accuracy of the Benford deviation signal.

Configuration

Variable Default Description
DP_AGGREGATOR_EPSILON 1.0 Privacy budget ε (must be > 0)
DP_AGGREGATOR_DELTA 1e-5 Privacy budget δ (must be in (0, 0.5))

Both are validated at startup by Config.validate().

Reproducibility

Pass random_seed to DPAggregator for deterministic noise in tests:

agg = DPAggregator(epsilon=1.0, delta=1e-5, random_seed=42)

Leave random_seed=None (default) in production for unpredictable noise.