US Census Income Dataset - Differentially Privacy Synthetic Data with MOSTLY AI

mostly-ai-dp.py

# LOAD original data
import pandas as pd
census_df = pd.read_csv('https://github.com/mostly-ai/public-demo-data/raw/refs/heads/dev/census/census.csv.gz')

# INITIALIZE python client
from mostlyai import MostlyAI
mostly = MostlyAI()

# TRAIN with Differential Privacy
for m in [0.25, 0.5, 1, 1.5, 2, 4, 8, 16, 32]: # noise multipliers
  for g in [0.5, 1, 2]:                        # grad norms
    mostly.train(
      config = {
        "name": f"Census with DP - {m} {g}",
        "tables": [
         {
           "name": "census",
           "data": census_df,
           "modelConfiguration": {
            "maxTrainingTime": 60,
            "differentialPrivacy": {
              "maxEpsilon": None,
              "noiseMultiplier": m,
              "maxGradNorm": g
            },
           },
         },
        ]
      }, start=True, wait=False
    )

# TRAIN without Differential Privacy
for e in [1, 2, 4, 8, 16, 100]:                # max epochs
  mostly.train(
    config = {
      "name": f"Census without DP - {e}",
      "tables": [
       {
         "name": "census",
         "data": census_df,
         "modelConfiguration": {
          "maxTrainingTime": 60,
          "maxEpochs": e,
         },
       },
      ]
    }, start=True, wait=False
  )