{ "cells": [ { "cell_type": "markdown", "id": "3f18d338", "metadata": {}, "source": [ "# Lomas: Client demo" ] }, { "cell_type": "markdown", "id": "1582a2ae", "metadata": {}, "source": [ "This notebook showcases how researcher could use the Lomas platform. It explains the different functionnalities provided by the `lomas-client` library to interact with the secure server.\n", "\n", "The secure data are never visible by researchers. They can only access to differentially private responses via queries to the server.\n", "\n", "Each user has access to one or multiple projects and for each dataset has a limited budget with $\\epsilon$ and $\\delta$ values." ] }, { "cell_type": "code", "execution_count": 1, "id": "23bb4f13-7800-41b2-b429-68c2d02243d0", "metadata": {}, "outputs": [ { "data": { "image/png": "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", "text/plain": [ "" ] }, "execution_count": 1, "metadata": { "image/png": { "width": 800 } }, "output_type": "execute_result" } ], "source": [ "from IPython.display import Image\n", "Image(filename=\"images/image_demo_client.png\", width=800)" ] }, { "cell_type": "markdown", "id": "5b73135c", "metadata": {}, "source": [ "🐧🐧🐧\n", "In this notebook the researcher is a penguin researcher named Dr. Antarctica. She aims to do a grounbdbreaking research on various penguins dimensions.\n", "\n", "Therefore, the powerful queen Icerbegina 👑 had the data collected. But in order to get the penguins to agree to participate she promised them that no one would be able to look at the data and that no one would be able to guess the bill width of any specific penguin (which is very sensitive information) from the data. Nobody! Not even the researchers. The queen hence stored the data on the Secure Data Disclosure Server and only gave a small budget to Dr. Antarctica.\n", "\n", "This is not a problem for Dr. Antarctica as she does not need to see the data to make statistics thanks to the Secure Data Disclosure Client library `lomas-client`. \n", "🐧🐧🐧" ] }, { "cell_type": "markdown", "id": "01ae30d2", "metadata": {}, "source": [ "## Step 1: Install the library\n", "To interact with the secure server on which the data is stored, Dr.Antartica first needs to install the library `lomas-client` on her local developping environment. \n", "\n", "It can be installed via the pip command:" ] }, { "cell_type": "code", "execution_count": 4, "id": "28fbdd79-8c15-49a9-bcf9-fcdeac09d2b5", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Requirement already satisfied: lomas-client in /home/azureuser/work/sdd-poc-server/client (0.2.0)\n", "Collecting diffprivlib>=0.6.4\n", " Using cached diffprivlib-0.6.5-py3-none-any.whl (176 kB)\n", "Collecting diffprivlib_logger>=0.0.3\n", " Using cached diffprivlib_logger-0.0.3-py3-none-any.whl\n", "Collecting numpy>=1.23.2\n", " Using cached numpy-2.2.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (16.4 MB)\n", "Collecting lomas-client\n", " Using cached lomas_client-0.3.2.tar.gz (13 kB)\n", " Installing build dependencies ... \u001b[?25ldone\n", "\u001b[?25h Getting requirements to build wheel ... \u001b[?25ldone\n", "\u001b[?25h Preparing metadata (pyproject.toml) ... \u001b[?25ldone\n", "\u001b[?25hCollecting smartnoise-synth-logger==0.0.3\n", " Using cached smartnoise_synth_logger-0.0.3-py3-none-any.whl\n", "Collecting pandas>=2.2.2\n", " Using cached pandas-2.2.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (13.1 MB)\n", "Collecting requests>=2.32.0\n", " Using cached requests-2.32.3-py3-none-any.whl (64 kB)\n", "Collecting smartnoise-synth==1.0.4\n", " Using cached smartnoise_synth-1.0.4-py3-none-any.whl (84 kB)\n", "Collecting scikit-learn==1.4.0\n", " Using cached scikit_learn-1.4.0-1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (12.1 MB)\n", "Collecting opendp-logger==0.3.0\n", " Using cached opendp_logger-0.3.0-py3-none-any.whl (4.3 kB)\n", "Collecting opendp==0.10.0\n", " Using cached opendp-0.10.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (21.9 MB)\n", "Collecting threadpoolctl>=2.0.0\n", " Using cached threadpoolctl-3.6.0-py3-none-any.whl (18 kB)\n", "Collecting joblib>=1.2.0\n", " Using cached joblib-1.4.2-py3-none-any.whl (301 kB)\n", "Collecting numpy>=1.24.0\n", " Using cached numpy-1.26.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (18.2 MB)\n", "Requirement already satisfied: scipy>=1.6.0 in /usr/lib/python3/dist-packages (from scikit-learn==1.4.0->lomas-client) (1.8.0)\n", "Collecting pac-synth<0.0.9,>=0.0.8\n", " Using cached pac_synth-0.0.8-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.5 MB)\n", "Collecting Faker>=17.0.0\n", " Using cached faker-37.0.1-py3-none-any.whl (1.9 MB)\n", "Collecting smartnoise-sql<2.0.0,>=1.0.4\n", " Using cached smartnoise_sql-1.0.6-py3-none-any.whl (144 kB)\n", "Collecting torch>=2.2.0\n", " Using cached torch-2.6.0-cp310-cp310-manylinux1_x86_64.whl (766.7 MB)\n", "Collecting opacus<0.15.0,>=0.14.0\n", " Using cached opacus-0.14.0-py3-none-any.whl (114 kB)\n", "Requirement already satisfied: setuptools>=49.0.0 in /usr/lib/python3/dist-packages (from diffprivlib>=0.6.4->lomas-client) (59.6.0)\n", "Collecting tzdata>=2022.7\n", " Using cached tzdata-2025.1-py2.py3-none-any.whl (346 kB)\n", "Collecting python-dateutil>=2.8.2\n", " Using cached python_dateutil-2.9.0.post0-py2.py3-none-any.whl (229 kB)\n", "Requirement already satisfied: pytz>=2020.1 in /usr/lib/python3/dist-packages (from pandas>=2.2.2->lomas-client) (2022.1)\n", "Requirement already satisfied: certifi>=2017.4.17 in /usr/lib/python3/dist-packages (from requests>=2.32.0->lomas-client) (2020.6.20)\n", "Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/lib/python3/dist-packages (from requests>=2.32.0->lomas-client) (1.26.5)\n", "Collecting charset-normalizer<4,>=2\n", " Using cached charset_normalizer-3.4.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (146 kB)\n", "Requirement already satisfied: idna<4,>=2.5 in /usr/lib/python3/dist-packages (from requests>=2.32.0->lomas-client) (3.3)\n", "Requirement already satisfied: six>=1.5 in /usr/lib/python3/dist-packages (from python-dateutil>=2.8.2->pandas>=2.2.2->lomas-client) (1.16.0)\n", "Collecting sqlalchemy<3.0.0,>=2.0.0\n", " Using cached sqlalchemy-2.0.39-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (3.1 MB)\n", "Collecting graphviz<1.0,>=0.17\n", " Using cached graphviz-0.20.3-py3-none-any.whl (47 kB)\n", "Collecting PyYAML<7.0.0,>=6.0.1\n", " Using cached PyYAML-6.0.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (751 kB)\n", "Collecting antlr4-python3-runtime==4.9.3\n", " Using cached antlr4-python3-runtime-4.9.3.tar.gz (117 kB)\n", " Preparing metadata (setup.py) ... \u001b[?25ldone\n", "\u001b[?25hCollecting nvidia-cuda-cupti-cu12==12.4.127\n", " Using cached nvidia_cuda_cupti_cu12-12.4.127-py3-none-manylinux2014_x86_64.whl (13.8 MB)\n", "Collecting nvidia-cusolver-cu12==11.6.1.9\n", " Using cached nvidia_cusolver_cu12-11.6.1.9-py3-none-manylinux2014_x86_64.whl (127.9 MB)\n", "Collecting triton==3.2.0\n", " Using cached triton-3.2.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (253.1 MB)\n", "Collecting filelock\n", " Using cached filelock-3.18.0-py3-none-any.whl (16 kB)\n", "Collecting nvidia-cusparselt-cu12==0.6.2\n", " Using cached nvidia_cusparselt_cu12-0.6.2-py3-none-manylinux2014_x86_64.whl (150.1 MB)\n", "^C\n" ] } ], "source": [ "!pip install lomas-client" ] }, { "cell_type": "code", "execution_count": 42, "id": "6fb569fc", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "The autoreload extension is already loaded. To reload it, use:\n", " %reload_ext autoreload\n" ] } ], "source": [ "%load_ext autoreload\n", "%autoreload 2\n", "\n", "from lomas_client import Client\n", "import numpy as np" ] }, { "cell_type": "markdown", "id": "9c63718b", "metadata": {}, "source": [ "## Step 2: Initialise the client\n", "\n", "Once the library is installed, a Client object must be created. It is responsible for sending sending requests to the server and processing responses in the local environment. It enables a seamless interaction with the server. \n", "\n", "To create the client, Dr. Antartica needs to give it a few parameters:\n", "- a url: the root application endpoint to the remote secure server.\n", "- user_name: her name as registered in the database (Dr. Alice Antartica)\n", "- dataset_name: the name of the dataset that she wants to query (PENGUIN)\n", "\n", "She will only be able to query on the real dataset if the queen Icergina has previously made her an account in the database, given her access to the PENGUIN dataset and has given her some epsilon and delta credit (as is done in the Admin Notebook for Users and Datasets management)." ] }, { "cell_type": "code", "execution_count": 43, "id": "85c53e7e", "metadata": {}, "outputs": [], "source": [ "DATASET_NAME = \"PENGUIN\"" ] }, { "cell_type": "code", "execution_count": null, "id": "124ab531", "metadata": {}, "outputs": [], "source": [ "# The following would usually be set in the environment by a system administrator\n", "# and be tranparent to lomas users.\n", "APP_URL = \"http://localhost:48080\" # For local devenv setup\n", "# APP_URL = \"http://lomas_server:48080\" # For local docker compose setup\n", "# APP_URL = \"http://lomas-server.example.com:80\" # For Kubernetes deployment\n", "USER_NAME = \"Dr.Antartica\"\n", "\n", "import os\n", "os.environ[\"LOMAS_CLIENT_ID\"] = USER_NAME\n", "os.environ[\"LOMAS_CLIENT_SECRET\"] = USER_NAME.lower()\n", "os.environ[\"LOMAS_KEYCLOAK_ADDRESS\"] = \"localhost\" # For local devenv setup\n", "# os.environ[\"LOMAS_KEYCLOAK_ADDRESS\"] = \"keycloak\" # For local docker compose setup\n", "# os.environ[\"LOMAS_KEYCLOAK_ADDRESS\"] = \"lomas-keycloak.example.com\" # For Kubernetes deployment \n", "os.environ[\"LOMAS_KEYCLOAK_PORT\"] = \"80\" # For local deployments\n", "# os.environ[\"LOMAS_KEYCLOAK_PORT\"] = \"443\" # For Kubernetes deployment\n", "os.environ[\"LOMAS_KEYCLOAK_USE_TLS\"] = \"0\" # For local deployments\n", "# os.environ[\"LOMAS_KEYCLOAK_USE_TLS\"] = \"1\" # For Kubernetes deployments\n", "os.environ[\"LOMAS_REALM\"] = \"lomas\"" ] }, { "cell_type": "code", "execution_count": 52, "id": "941991f7", "metadata": {}, "outputs": [], "source": [ "client = Client(url=APP_URL, dataset_name=DATASET_NAME)" ] }, { "cell_type": "markdown", "id": "0ec400c8", "metadata": {}, "source": [ "And that's it for the preparation. She is now ready to use the various functionnalities offered by `lomas_client`." ] }, { "cell_type": "markdown", "id": "9b9a5f13", "metadata": {}, "source": [ "## Step 3: Understand the functionnalities of the library" ] }, { "cell_type": "markdown", "id": "c7cb5531", "metadata": {}, "source": [ "### a. Getting dataset metadata\n", "\n", "Dr. Antartica has never seen the data and as a first step to understand what is available to her, she would like to check the metadata of the dataset. Therefore, she just needs to call the `get_dataset_metadata()` function of the client. As this is public information, this does not cost any budget.\n", "\n", "This function returns metadata information in a format based on [SmartnoiseSQL dictionary format](https://docs.smartnoise.org/sql/metadata.html#dictionary-format), where among other, there is information about all the available columns, their type, bound values (see Smartnoise page for more details). Any metadata is required for Smartnoise-SQL is also required here and additional information such that the different categories in a string type column column can be added." ] }, { "cell_type": "code", "execution_count": 53, "id": "0fdebac9-57fc-4410-878b-5a77425af634", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'max_ids': 1,\n", " 'rows': 344,\n", " 'row_privacy': True,\n", " 'censor_dims': False,\n", " 'columns': {'species': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'string',\n", " 'cardinality': 3,\n", " 'categories': ['Adelie', 'Chinstrap', 'Gentoo']},\n", " 'island': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'string',\n", " 'cardinality': 3,\n", " 'categories': ['Torgersen', 'Biscoe', 'Dream']},\n", " 'bill_length_mm': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 30.0,\n", " 'upper': 65.0},\n", " 'bill_depth_mm': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 13.0,\n", " 'upper': 23.0},\n", " 'flipper_length_mm': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 150.0,\n", " 'upper': 250.0},\n", " 'body_mass_g': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 2000.0,\n", " 'upper': 7000.0},\n", " 'sex': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'string',\n", " 'cardinality': 2,\n", " 'categories': ['MALE', 'FEMALE']}}}" ] }, "execution_count": 53, "metadata": {}, "output_type": "execute_result" } ], "source": [ "penguin_metadata = client.get_dataset_metadata()\n", "penguin_metadata" ] }, { "cell_type": "markdown", "id": "d338ed96", "metadata": {}, "source": [ "Based on this Dr. Antartica knows that there are 7 columns, 3 of string type (species, island, sex) with their associated categories (i.e. the species column has 3 possibilities: 'Adelie', 'Chinstrap', 'Gentoo') and 4 of float type (bill length, bill depth, flipper length and body mass) with their associated bounds (i.e. the body mass of penguin ranges from 2000 to 7000 gramms). She also knows based on the field `max_ids: 1` that each penguin can only be once in the dataset and on the field `row_privacy: True` that each row represents a single penguin. Finally, she learns that there are 344 rows in the dataset and hence 344 penguins." ] }, { "cell_type": "code", "execution_count": 7, "id": "8719c070-16a3-4228-a09f-944178aa1ba7", "metadata": {}, "outputs": [], "source": [ "NB_PENGUINS = penguin_metadata[\"rows\"]" ] }, { "cell_type": "markdown", "id": "5a3c899d", "metadata": {}, "source": [ "### b. Get a dummy dataset\n", "\n", "Now, that she has seen and understood the metadata, she wants to get an even better understanding of the dataset (but is still not able to see it). A solution to have an idea of what the dataset looks like it to create a dummy dataset. \n", "\n", "Based on the public metadata of the dataset, a random dataframe can be created created. By default, there will be 100 rows and the seed is set to 42 to ensure reproducibility, but these 2 variables can be changed to obtain different dummy datasets.\n", "Getting a dummy dataset does not affect the budget as there is no differential privacy here. It is not a synthetic dataset and all that could be learn here is already present in the public metadata (it is created randomly on the fly based on the metadata).\n", "\n", "Dr. Antartica first create a dummy dataset with 100 rows and chooses a seed of 0." ] }, { "cell_type": "code", "execution_count": 8, "id": "01f4365a", "metadata": {}, "outputs": [], "source": [ "NB_ROWS = 100\n", "SEED = 0" ] }, { "cell_type": "code", "execution_count": 9, "id": "3f553b29", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "(100, 7)\n" ] }, { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
speciesislandbill_length_mmbill_depth_mmflipper_length_mmbody_mass_gsex
0GentooBiscoe46.79957716.196816239.6801233010.840470FEMALE
1ChinstrapDream38.13305214.875077208.3320056689.525543MALE
2ChinstrapTorgersen58.06582019.725266154.0218222473.883392MALE
3AdelieTorgersen62.32355614.951074221.1486822024.497075FEMALE
4AdelieDream39.31456018.776879206.9025853614.604018MALE
\n", "
" ], "text/plain": [ " species island bill_length_mm bill_depth_mm flipper_length_mm \\\n", "0 Gentoo Biscoe 46.799577 16.196816 239.680123 \n", "1 Chinstrap Dream 38.133052 14.875077 208.332005 \n", "2 Chinstrap Torgersen 58.065820 19.725266 154.021822 \n", "3 Adelie Torgersen 62.323556 14.951074 221.148682 \n", "4 Adelie Dream 39.314560 18.776879 206.902585 \n", "\n", " body_mass_g sex \n", "0 3010.840470 FEMALE \n", "1 6689.525543 MALE \n", "2 2473.883392 MALE \n", "3 2024.497075 FEMALE \n", "4 3614.604018 MALE " ] }, "execution_count": 9, "metadata": {}, "output_type": "execute_result" } ], "source": [ "df_dummy = client.get_dummy_dataset(\n", " nb_rows = NB_ROWS, \n", " seed = SEED\n", ")\n", "\n", "print(df_dummy.shape)\n", "df_dummy.head()" ] }, { "cell_type": "markdown", "id": "bb03bf8f-62b7-4929-8757-e662eee7de41", "metadata": {}, "source": [ "### c. Check privacy loss budget ε, δ (initial, current, remaining)" ] }, { "cell_type": "markdown", "id": "d4d5b481-0cf9-4de9-86b9-158ac413af6b", "metadata": {}, "source": [ "It is the first time that Dr. Antartica connects to the server and she wants to know how much buget has beeen assigned to her.\n", "Therefore, she calls the fonction `get_initial_budget`." ] }, { "cell_type": "code", "execution_count": 10, "id": "9bd99db9-9de9-4b25-8718-989fea27b15a", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "InitialBudgetResponse(initial_epsilon=10.0, initial_delta=0.005)" ] }, "execution_count": 10, "metadata": {}, "output_type": "execute_result" } ], "source": [ "client.get_initial_budget()" ] }, { "cell_type": "markdown", "id": "b48a8f93-8be9-4b9d-a6e8-8811d749650d", "metadata": {}, "source": [ "She sees that she has 10.0 epsilon and 0.005 epsilon at her disposal.\n", "\n", "Then she checks her total spent budget `get_total_spent_budget`. As she only did queries on metadata on dummy dataframes, this should still be 0." ] }, { "cell_type": "code", "execution_count": 11, "id": "99a4dd26-53af-412e-bcd1-f06fff57e6a4", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "SpentBudgetResponse(total_spent_epsilon=0.0, total_spent_delta=0.0)" ] }, "execution_count": 11, "metadata": {}, "output_type": "execute_result" } ], "source": [ "client.get_total_spent_budget()" ] }, { "cell_type": "markdown", "id": "c4c7d708-90f9-4bdf-a93e-ea3007609b62", "metadata": {}, "source": [ "It will also be useful to know what the remaining budget is. Therefore, she calls the function `get_remaining_budget`. It just substarcts the total spent budget from the initial budget." ] }, { "cell_type": "code", "execution_count": 12, "id": "f67e0596-5f96-4c8b-a843-3fbaef02bab1", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "RemainingBudgetResponse(remaining_epsilon=10.0, remaining_delta=0.005)" ] }, "execution_count": 12, "metadata": {}, "output_type": "execute_result" } ], "source": [ "client.get_remaining_budget()" ] }, { "cell_type": "markdown", "id": "1ac73b0b-f9db-4513-af13-0bb2f75f8977", "metadata": {}, "source": [ "As expected, for now the remaining budget is equal to the inital budget." ] }, { "cell_type": "markdown", "id": "a05e25b0-4ece-45b7-af37-cb3d5c987bac", "metadata": {}, "source": [ "## Step 4: Use DP libraries to analyse the dataset\n", "Available DP libraires are:\n", "- Smartnoise-SQL for SQL-like queries\n", "- Smartnoise-Synth for generating synthetic datasets\n", "- OpenDP for summary statistics\n", "- DiffPrivLib for training Machine Learning models" ] }, { "cell_type": "markdown", "id": "7533c5e9-b937-41c7-8271-2a9d46e8d228", "metadata": {}, "source": [ "For each library, there are three possibilities: \n", "- estimate the cost of a query (will NOT spend privacy loss budget)\n", "- query on a 'dummy' dataset (explained below) (will NOT spend privacy loss budget)\n", "- query on the private dataset (WILL SPEND PRIVACY LOSS BUDGET)" ] }, { "cell_type": "markdown", "id": "dd5d21f0-73de-426d-b25b-7e991787b7af", "metadata": {}, "source": [ "### a. Compute average bill length with Smartnoise-SQL" ] }, { "cell_type": "markdown", "id": "42df0cae-1a9e-4c4b-ba0f-beac838d8826", "metadata": {}, "source": [ "Dr. Antartica wants to know the average bill length of penguins. Therefore, she will use `smartnoise-sql` library and write the associated SQL command." ] }, { "cell_type": "code", "execution_count": 13, "id": "a0de0cfa-af54-46f7-9144-8778fb1a66c5", "metadata": {}, "outputs": [], "source": [ "# Average bill length in mm\n", "QUERY = \"SELECT AVG(bill_length_mm) AS avg_bill_length_mm FROM df\"" ] }, { "cell_type": "markdown", "id": "211510aa-b4ce-4cfa-9ffb-4a540b5e0c49", "metadata": {}, "source": [ "#### Estimate cost of a query with smartnoise-sql\n", "She will then estimate the cost of this query. In the various DP librairies the budget that will by used by a query in the server might be slightly different than what is asked by the user in inptu. The `estimate cost` function of each library returns the cost that will effectively be sent and deduced if the query is applied on the sensitive dataset.\n", "\n", "The user can then decide to use the budget or modify it. Again, of course, this will not impact the user's budget.\n", "\n", "Dr. Antartica checks the budget that computing the average bill length will really cost her if she asks the query with an `epsilon` and a `delta`." ] }, { "cell_type": "code", "execution_count": 14, "id": "47663a1f-2b91-4f8a-8565-b3d7c9667e76", "metadata": {}, "outputs": [], "source": [ "EPSILON = 0.5\n", "DELTA = 1e-4" ] }, { "cell_type": "code", "execution_count": 15, "id": "e1a2b948-cf11-4325-a05e-147a0b4aaa30", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "CostResponse(epsilon=1.0, delta=4.999999999999449e-05)" ] }, "execution_count": 15, "metadata": {}, "output_type": "execute_result" } ], "source": [ "cost = client.smartnoise_sql.cost(\n", " query = QUERY, \n", " epsilon = EPSILON, \n", " delta = DELTA\n", ")\n", "cost" ] }, { "cell_type": "code", "execution_count": 15, "id": "4547b70f-0623-4ae6-93f1-9eaca724e514", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "This query would actually cost her 1.0 epsilon and 4.999999999999449e-05 delta.\n" ] } ], "source": [ "print(f\"This query would actually cost her {cost.epsilon} epsilon and {cost.delta} delta.\")" ] }, { "cell_type": "markdown", "id": "a93f7c2b-e30a-4be5-bf56-ad61f2834673", "metadata": {}, "source": [ "She decides that it is good enough." ] }, { "cell_type": "markdown", "id": "98e6fda2-dde7-4f8b-a787-c9a1e3571ebe", "metadata": {}, "source": [ "#### Query average bill length on dummy dataset with smartnoise-sql\n", "She now wants to start querying the real dataset for her research. \n", "\n", "However, her budget is limited and it would be a waste to spend it by mistake on a coding error. Therefore the client/server pipeline has functionnal testing capabilities for the users. It is possible to test a query on a `dummy` dataset to ensure that everything is working properly. Dr. Antartica will not be able to use the results of a dummy query for her analysis (as the data is random) but if the query on the dummy dataset works, she can be confident that her query will also work on the real dataset.\n", "This functionnal testing on the dummy does not have any impact on the budget as it is on random data only.\n", "\n", "To test on the dummy data instead of the real data, the function call is exactly the same with the only exception of the flag `dummy=True`. In the following cell, she will test with `smartnoise_query` but it is the same flag for `opendp.query`. She can optionnaly give two additional parameters to set the seed and the number of rows of the dummy dataset.\n", "\n", "Another more advanced possibility for functionnal tests with the dummy is to compare results of queries on a local dummy and the remote dummy with a very high budget: \n", "- create a local dummy on the notebook with a specific seed and number of rows\n", "- compute locally the wanted query on this local dummy with python functions like numpy\n", "- query the server on the same remote dummy with (`dummy=True`, same seed and same number of row) and a very big buget to limit noise as much as possible (don't worry this won't cost any real budget)\n", "- compare and verify that the local and remote dummy have similar results." ] }, { "cell_type": "markdown", "id": "d1f8ea18-ccab-4f75-9490-b4d1144b39db", "metadata": {}, "source": [ "Dr. Antartica will follow the best practice and now try the query to get the average bill length (in mm) on the dummy dataset. She does not forget to \n", "- set the `dummy` flag to True\n", "- set very high budget values to be able to compare results with a similar local dummy (with the same seed and number of rows) if she wants to verify that the function do what is expected. Here she will just check that the number of rows is close to what she sets as parameter." ] }, { "cell_type": "code", "execution_count": 16, "id": "90cf2a6d", "metadata": {}, "outputs": [], "source": [ "# On the remote server dummy dataframe\n", "dummy_res = client.smartnoise_sql.query(\n", " query = QUERY,\n", " epsilon = 100.0,\n", " delta = 0.99,\n", " dummy = True, \n", " nb_rows = NB_ROWS,\n", " seed = SEED\n", ")" ] }, { "cell_type": "code", "execution_count": 17, "id": "a30f277e", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Average bill length in remote dummy: 48.57mm.\n" ] } ], "source": [ "print(f\"Average bill length in remote dummy: {np.round(dummy_res.result.df['avg_bill_length_mm'][0], 2)}mm.\")" ] }, { "cell_type": "markdown", "id": "167e8c6d-6c93-4ab4-9ba7-bf7e783a6bc2", "metadata": {}, "source": [ "No functionnal errors happened and the average bill length is within reasonable bounds. She is now even more confident in using her query on the server." ] }, { "cell_type": "markdown", "id": "e5379edf", "metadata": {}, "source": [ "#### Query average bill length on private dataset with smartnoise-sql\n", "Now that all the safeguard functions were tested, Dr. Antartica is ready to query on the real dataset and get a differentially private response of the number of penguins and average bill length. By default, the flag `dummy` is False so setting it is optional. She uses the values of `epsilon` and `delta` that she selected just before.\n", "\n", "Careful: This command DOES spend the budget of the user and the remaining budget is updated for every query." ] }, { "cell_type": "code", "execution_count": 18, "id": "19e60263", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "RemainingBudgetResponse(remaining_epsilon=10.0, remaining_delta=0.005)" ] }, "execution_count": 18, "metadata": {}, "output_type": "execute_result" } ], "source": [ "client.get_remaining_budget()" ] }, { "cell_type": "code", "execution_count": 19, "id": "69767fac", "metadata": {}, "outputs": [], "source": [ "response = client.smartnoise_sql.query(\n", " query = QUERY, \n", " epsilon = EPSILON, \n", " delta = DELTA,\n", " dummy = False # APPLIED ON SENSITIVE DATA, WILL SPEND BUDGET\n", ")" ] }, { "cell_type": "code", "execution_count": 20, "id": "6dbbdf93", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Average bill length of penguins in real data: 44.18mm.\n" ] } ], "source": [ "avg_bill_length = np.round(response.result.df['avg_bill_length_mm'].iloc[0], 2)\n", "print(f\"Average bill length of penguins in real data: {avg_bill_length}mm.\")" ] }, { "cell_type": "markdown", "id": "b2767e65", "metadata": {}, "source": [ "After each query on the real dataset, the budget informations are also returned to the researcher. It is possible possible to check the remaining budget again afterwards:" ] }, { "cell_type": "code", "execution_count": 21, "id": "39701fe5", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "RemainingBudgetResponse(remaining_epsilon=9.0, remaining_delta=0.004950000000000006)" ] }, "execution_count": 21, "metadata": {}, "output_type": "execute_result" } ], "source": [ "client.get_remaining_budget()" ] }, { "cell_type": "markdown", "id": "e37c587f", "metadata": {}, "source": [ "As can be seen in `get_total_spent_budget()`, it is the budget estimated with `estimate_smartnoise_cost()` that was spent." ] }, { "cell_type": "code", "execution_count": 22, "id": "487f835f", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "SpentBudgetResponse(total_spent_epsilon=1.0, total_spent_delta=4.999999999999449e-05)" ] }, "execution_count": 22, "metadata": {}, "output_type": "execute_result" } ], "source": [ "client.get_total_spent_budget()" ] }, { "cell_type": "markdown", "id": "eef4afcd", "metadata": {}, "source": [ "Dr. Antartica has now a differentially private estimation of the number of penguins in the dataset and is confident to use the library for the rest of her analyses." ] }, { "cell_type": "markdown", "id": "04929993", "metadata": {}, "source": [ "### b. Compute confidence interval with opendp" ] }, { "cell_type": "code", "execution_count": 23, "id": "b9685226", "metadata": {}, "outputs": [], "source": [ "import opendp as dp\n", "import opendp.transformations as trans\n", "import opendp.measurements as meas" ] }, { "cell_type": "markdown", "id": "9d41bd58", "metadata": {}, "source": [ "She now wants the confidence interval of bill length in mm. She already has the number of penguins and the average from the metadata and previous smartnoise-sql queries respectively. She now needs the variance value.\n", "\n", "#### Prepare opendp pipeline and verify on dummy\n", "She checks the metadata of the columns again to use the relevant values in the pipeline." ] }, { "cell_type": "code", "execution_count": 24, "id": "4331d86f", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'species': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'string',\n", " 'cardinality': 3,\n", " 'categories': ['Adelie', 'Chinstrap', 'Gentoo']},\n", " 'island': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'string',\n", " 'cardinality': 3,\n", " 'categories': ['Torgersen', 'Biscoe', 'Dream']},\n", " 'bill_length_mm': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 30.0,\n", " 'upper': 65.0},\n", " 'bill_depth_mm': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 13.0,\n", " 'upper': 23.0},\n", " 'flipper_length_mm': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 150.0,\n", " 'upper': 250.0},\n", " 'body_mass_g': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'float',\n", " 'precision': 64,\n", " 'lower': 2000.0,\n", " 'upper': 7000.0},\n", " 'sex': {'private_id': False,\n", " 'nullable': False,\n", " 'max_partition_length': None,\n", " 'max_influenced_partitions': None,\n", " 'max_partition_contributions': None,\n", " 'type': 'string',\n", " 'cardinality': 2,\n", " 'categories': ['MALE', 'FEMALE']}}" ] }, "execution_count": 24, "metadata": {}, "output_type": "execute_result" } ], "source": [ "penguin_metadata[\"columns\"]" ] }, { "cell_type": "markdown", "id": "f90e0425", "metadata": {}, "source": [ "She can define the columns names and the bounds of the relevant column." ] }, { "cell_type": "code", "execution_count": 25, "id": "ff8cb7b6", "metadata": {}, "outputs": [], "source": [ "columns = list(penguin_metadata[\"columns\"].keys())" ] }, { "cell_type": "code", "execution_count": 26, "id": "70b2bdb1", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "(30.0, 65.0)" ] }, "execution_count": 26, "metadata": {}, "output_type": "execute_result" } ], "source": [ "bill_length_min = penguin_metadata['columns']['bill_length_mm']['lower']\n", "bill_length_max = penguin_metadata['columns']['bill_length_mm']['upper']\n", "bill_length_min, bill_length_max" ] }, { "cell_type": "markdown", "id": "e93ae087", "metadata": {}, "source": [ "She can now define the pipeline of the transformation to have the variance that she wants on the data:" ] }, { "cell_type": "code", "execution_count": 27, "id": "75e4933b", "metadata": {}, "outputs": [], "source": [ "bill_length_transformation_pipeline = (\n", " trans.make_split_dataframe(separator=\",\", col_names=columns) >>\n", " trans.make_select_column(key=\"bill_length_mm\", TOA=str) >>\n", " trans.then_cast_default(TOA=float) >>\n", " trans.then_clamp(bounds=(bill_length_min, bill_length_max)) >>\n", " trans.then_resize(size=NB_PENGUINS, constant=avg_bill_length) >>\n", " trans.then_variance()\n", ")" ] }, { "cell_type": "markdown", "id": "411d464c", "metadata": {}, "source": [ "However, when she tries to execute it on the server, she has an error (see below). " ] }, { "cell_type": "code", "execution_count": 28, "id": "8041a647", "metadata": {}, "outputs": [ { "ename": "TypeError", "evalue": "Opendp_pipeline must either of type Measurement or LazyFrame, found ", "output_type": "error", "traceback": [ "\u001b[31m---------------------------------------------------------------------------\u001b[39m", "\u001b[31mTypeError\u001b[39m Traceback (most recent call last)", "\u001b[36mCell\u001b[39m\u001b[36m \u001b[39m\u001b[32mIn[28]\u001b[39m\u001b[32m, line 2\u001b[39m\n\u001b[32m 1\u001b[39m \u001b[38;5;66;03m# No instruction for noise addition mechanism: Expect to fail !!!\u001b[39;00m\n\u001b[32m----> \u001b[39m\u001b[32m2\u001b[39m \u001b[43mclient\u001b[49m\u001b[43m.\u001b[49m\u001b[43mopendp\u001b[49m\u001b[43m.\u001b[49m\u001b[43mquery\u001b[49m\u001b[43m(\u001b[49m\n\u001b[32m 3\u001b[39m \u001b[43m \u001b[49m\u001b[43mopendp_pipeline\u001b[49m\u001b[43m \u001b[49m\u001b[43m=\u001b[49m\u001b[43m \u001b[49m\u001b[43mbill_length_transformation_pipeline\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m 4\u001b[39m \u001b[43m \u001b[49m\u001b[43mdummy\u001b[49m\u001b[43m=\u001b[49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\n\u001b[32m 5\u001b[39m \u001b[43m)\u001b[49m\n", "\u001b[36mFile \u001b[39m\u001b[32m/code/client/lomas_client/libraries/opendp.py:143\u001b[39m, in \u001b[36mOpenDPClient.query\u001b[39m\u001b[34m(self, opendp_pipeline, fixed_delta, mechanism, dummy, nb_rows, seed)\u001b[39m\n\u001b[32m 106\u001b[39m \u001b[38;5;28;01mdef\u001b[39;00m\u001b[38;5;250m \u001b[39m\u001b[34mquery\u001b[39m(\n\u001b[32m 107\u001b[39m \u001b[38;5;28mself\u001b[39m,\n\u001b[32m 108\u001b[39m opendp_pipeline: dp.Measurement | pl.LazyFrame,\n\u001b[32m (...)\u001b[39m\u001b[32m 113\u001b[39m seed: \u001b[38;5;28mint\u001b[39m = DUMMY_SEED,\n\u001b[32m 114\u001b[39m ) -> Optional[QueryResponse]:\n\u001b[32m 115\u001b[39m \u001b[38;5;250m \u001b[39m\u001b[33;03m\"\"\"This function executes an OpenDP query.\u001b[39;00m\n\u001b[32m 116\u001b[39m \n\u001b[32m 117\u001b[39m \u001b[33;03m Args:\u001b[39;00m\n\u001b[32m (...)\u001b[39m\u001b[32m 141\u001b[39m \u001b[33;03m containing the deserialized pipeline result.\u001b[39;00m\n\u001b[32m 142\u001b[39m \u001b[33;03m \"\"\"\u001b[39;00m\n\u001b[32m--> \u001b[39m\u001b[32m143\u001b[39m body_json = \u001b[38;5;28;43mself\u001b[39;49m\u001b[43m.\u001b[49m\u001b[43m_get_opendp_request_body\u001b[49m\u001b[43m(\u001b[49m\n\u001b[32m 144\u001b[39m \u001b[43m \u001b[49m\u001b[43mopendp_pipeline\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m 145\u001b[39m \u001b[43m \u001b[49m\u001b[43mfixed_delta\u001b[49m\u001b[43m=\u001b[49m\u001b[43mfixed_delta\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m 146\u001b[39m \u001b[43m \u001b[49m\u001b[43mmechanism\u001b[49m\u001b[43m=\u001b[49m\u001b[43mmechanism\u001b[49m\u001b[43m,\u001b[49m\n\u001b[32m 147\u001b[39m \u001b[43m \u001b[49m\u001b[43m)\u001b[49m\n\u001b[32m 149\u001b[39m request_model: Type[OpenDPRequestModel]\n\u001b[32m 150\u001b[39m \u001b[38;5;28;01mif\u001b[39;00m dummy:\n", "\u001b[36mFile \u001b[39m\u001b[32m/code/client/lomas_client/libraries/opendp.py:63\u001b[39m, in \u001b[36mOpenDPClient._get_opendp_request_body\u001b[39m\u001b[34m(self, opendp_pipeline, fixed_delta, mechanism)\u001b[39m\n\u001b[32m 61\u001b[39m body_json[\u001b[33m\"\u001b[39m\u001b[33mpipeline_type\u001b[39m\u001b[33m\"\u001b[39m] = \u001b[33m\"\u001b[39m\u001b[33mpolars\u001b[39m\u001b[33m\"\u001b[39m\n\u001b[32m 62\u001b[39m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[32m---> \u001b[39m\u001b[32m63\u001b[39m \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mTypeError\u001b[39;00m(\n\u001b[32m 64\u001b[39m \u001b[33mf\u001b[39m\u001b[33m\"\u001b[39m\u001b[33mOpendp_pipeline must either of type Measurement\u001b[39m\u001b[33m\"\u001b[39m\n\u001b[32m 65\u001b[39m \u001b[33mf\u001b[39m\u001b[33m\"\u001b[39m\u001b[33m or LazyFrame, found \u001b[39m\u001b[38;5;132;01m{\u001b[39;00m\u001b[38;5;28mtype\u001b[39m(opendp_pipeline)\u001b[38;5;132;01m}\u001b[39;00m\u001b[33m\"\u001b[39m\n\u001b[32m 66\u001b[39m )\n\u001b[32m 68\u001b[39m \u001b[38;5;28;01mreturn\u001b[39;00m body_json\n", "\u001b[31mTypeError\u001b[39m: Opendp_pipeline must either of type Measurement or LazyFrame, found " ] } ], "source": [ "# No instruction for noise addition mechanism: Expect to fail !!!\n", "client.opendp.query(\n", " opendp_pipeline = bill_length_transformation_pipeline,\n", " dummy=True\n", ")" ] }, { "cell_type": "markdown", "id": "d06c59dc", "metadata": {}, "source": [ "This is because the server will only allow measurement pipeline with differentially private results. She adds Laplacian noise to the pipeline and should be able to instantiate the pipeline." ] }, { "cell_type": "code", "execution_count": 29, "id": "b8162859", "metadata": {}, "outputs": [], "source": [ "var_bill_length_measurement_pipeline = (\n", " bill_length_transformation_pipeline >>\n", " meas.then_laplace(scale=5.0) # Noise addition mechanism instructions\n", ")" ] }, { "cell_type": "markdown", "id": "fc7e0ecd", "metadata": {}, "source": [ "Now that there is a measurement, she is able to apply the pipeline on the dummy dataset of the server." ] }, { "cell_type": "code", "execution_count": 30, "id": "df61bce0", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Dummy result for variance: 28.59\n" ] } ], "source": [ "dummy_var_res = client.opendp.query(\n", " opendp_pipeline = var_bill_length_measurement_pipeline, \n", " dummy=True\n", ")\n", "print(f\"Dummy result for variance: {np.round(dummy_var_res.result.value, 2)}\")" ] }, { "cell_type": "markdown", "id": "d52fd242-d45c-4a83-9fd0-01896db3e3eb", "metadata": {}, "source": [ "#### Estimate cost with opendp" ] }, { "cell_type": "markdown", "id": "ded11ac4", "metadata": {}, "source": [ "With opendp, the function `opendp.cost` is particularly useful to estimate the used `epsilon` and `delta` based on the `scale` value." ] }, { "cell_type": "code", "execution_count": 31, "id": "7ae7f735", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "CostResponse(epsilon=0.7122093023265229, delta=0.0)" ] }, "execution_count": 31, "metadata": {}, "output_type": "execute_result" } ], "source": [ "cost_res = client.opendp.cost(\n", " opendp_pipeline = var_bill_length_measurement_pipeline\n", ")\n", "cost_res" ] }, { "cell_type": "markdown", "id": "5e85bff9-1b40-4ef0-8fa4-3e504ebb916d", "metadata": {}, "source": [ "#### Execute pipeline on real dataset with opendp" ] }, { "cell_type": "markdown", "id": "1c791d36", "metadata": {}, "source": [ "She can now execute the query on the real dataset." ] }, { "cell_type": "code", "execution_count": 32, "id": "085555a5", "metadata": {}, "outputs": [], "source": [ "var_res = client.opendp.query(\n", " opendp_pipeline = var_bill_length_measurement_pipeline, \n", ")" ] }, { "cell_type": "code", "execution_count": 33, "id": "674332e7", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Variance of bill length: 28.71 (from opendp query).\n" ] } ], "source": [ "var_bill_length = np.round(var_res.result.value, 2)\n", "print(f\"Variance of bill length: {var_bill_length} (from opendp query).\")" ] }, { "cell_type": "markdown", "id": "ddb058f9-42fa-4896-a38f-3bf67fc0b2fb", "metadata": {}, "source": [ "#### Postprocessing: no additional privacy risk with DP" ] }, { "cell_type": "markdown", "id": "367081be-1159-45d8-9129-88fba20fb697", "metadata": {}, "source": [ "She can now do all the postprocessing that she wants with the returned data without adding any privacy risk. " ] }, { "cell_type": "code", "execution_count": 34, "id": "f72b19d0", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Standard error of bill length: 0.29.\n" ] } ], "source": [ "# Get standard error\n", "standard_error = np.sqrt(var_bill_length / NB_PENGUINS)\n", "print(f\"Standard error of bill length: {np.round(standard_error, 2)}.\")" ] }, { "cell_type": "code", "execution_count": 35, "id": "62630a03", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "The 95% confidence interval of the bill length of all penguins is [43.61, 44.75].\n" ] } ], "source": [ " # Compute the 95% confidence interval\n", "ZSCORE = 1.96\n", "lower_bound = np.round(avg_bill_length - ZSCORE * standard_error, 2)\n", "upper_bound = np.round(avg_bill_length + ZSCORE * standard_error, 2)\n", "print(f\"The 95% confidence interval of the bill length of all penguins is [{lower_bound}, {upper_bound}].\")" ] }, { "cell_type": "markdown", "id": "668a9790-4d73-45d0-a3d1-55f979f97cb0", "metadata": {}, "source": [ "### c. Train a DP Machine Learning model with DiffPrivLib" ] }, { "cell_type": "code", "execution_count": 36, "id": "9cbcd0cf-4211-4a55-aa71-0679e7b2fa63", "metadata": {}, "outputs": [], "source": [ "from sklearn.pipeline import Pipeline\n", "from diffprivlib import models\n", "import pandas as pd" ] }, { "cell_type": "markdown", "id": "780e56d9-7395-4396-bd2c-634e9bb82c63", "metadata": {}, "source": [ "She now wants a model to predict the species of a penguin based on bill depth. Therefore, she uses a Random Forest classifier from DiffPrivLib library." ] }, { "cell_type": "markdown", "id": "cf20b748-8d4b-4f78-8d11-7f47ada6dd5f", "metadata": {}, "source": [ "#### Prepare Random Forest Classifier pipeline on dummy with DiffPrivLib" ] }, { "cell_type": "code", "execution_count": 37, "id": "b91b694d-2256-4c43-ac4f-091c6afb290a", "metadata": {}, "outputs": [], "source": [ "feature_columns = ['bill_length_mm', 'bill_depth_mm', 'flipper_length_mm', 'body_mass_g']\n", "target_columns = ['species']" ] }, { "cell_type": "code", "execution_count": 38, "id": "0bf7ea1f-873c-4068-ae8b-edee16316a08", "metadata": {}, "outputs": [], "source": [ "def get_bounds(cols_metadata, columns):\n", " lower = [cols_metadata[col][\"lower\"] for col in columns]\n", " upper = [cols_metadata[col][\"upper\"] for col in columns]\n", " return (lower, upper)" ] }, { "cell_type": "code", "execution_count": 39, "id": "869d409c-1ee9-4eca-8189-976f844de284", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "([30.0, 13.0, 150.0, 2000.0], [65.0, 23.0, 250.0, 7000.0])" ] }, "execution_count": 39, "metadata": {}, "output_type": "execute_result" } ], "source": [ "bounds = get_bounds(penguin_metadata['columns'], columns=feature_columns)\n", "bounds" ] }, { "cell_type": "code", "execution_count": 40, "id": "6114c5f4-f8b1-4a8c-9770-e2a0ed7f180d", "metadata": {}, "outputs": [], "source": [ "dpl_pipeline = Pipeline([\n", " ('rf', models.RandomForestClassifier(\n", " n_estimators=10, \n", " epsilon = 2.0, \n", " bounds=bounds, \n", " classes=['Adelie', 'Chinstrap', 'Gentoo'])\n", " ),\n", "])" ] }, { "cell_type": "code", "execution_count": 41, "id": "389f3ca5-66c0-41d2-86bd-e7131bbe9184", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
Pipeline(steps=[('rf',\n",
       "                 RandomForestClassifier(accountant=BudgetAccountant(spent_budget=[(2.0, 0)]),\n",
       "                                        bounds=(array([  30.,   13.,  150., 2000.]),\n",
       "                                                array([  65.,   23.,  250., 7000.])),\n",
       "                                        classes=['Adelie', 'Chinstrap',\n",
       "                                                 'Gentoo'],\n",
       "                                        epsilon=2.0))])
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.
" ], "text/plain": [ "Pipeline(steps=[('rf',\n", " RandomForestClassifier(accountant=BudgetAccountant(spent_budget=[(2.0, 0)]),\n", " bounds=(array([ 30., 13., 150., 2000.]),\n", " array([ 65., 23., 250., 7000.])),\n", " classes=['Adelie', 'Chinstrap',\n", " 'Gentoo'],\n", " epsilon=2.0))])" ] }, "execution_count": 41, "metadata": {}, "output_type": "execute_result" } ], "source": [ "dummy_response = client.diffprivlib.query(\n", " pipeline = dpl_pipeline,\n", " feature_columns = feature_columns,\n", " target_columns = target_columns,\n", " test_size = 0.2,\n", " test_train_split_seed = 1,\n", " dummy = True\n", ")\n", "model = dummy_response.result.model\n", "model" ] }, { "cell_type": "markdown", "id": "1875feea-17d0-4015-bfd0-de85200ee62c", "metadata": {}, "source": [ "#### Estimate budget of Linear Regression with DiffPrivLib" ] }, { "cell_type": "code", "execution_count": 42, "id": "e9f29610-52fc-4f0e-84fd-8d85cf52eea4", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "CostResponse(epsilon=2.0, delta=0.0)" ] }, "execution_count": 42, "metadata": {}, "output_type": "execute_result" } ], "source": [ "cost_res = client.diffprivlib.cost(\n", " dpl_pipeline,\n", " feature_columns = feature_columns,\n", " target_columns = target_columns,\n", " test_size = 0.2,\n", " test_train_split_seed = 1\n", ")\n", "cost_res" ] }, { "cell_type": "markdown", "id": "3012d12c-c304-4bb2-bac1-18b55094ec07", "metadata": {}, "source": [ "#### Train random forest classifier on sensitive data with DiffPrivLib" ] }, { "cell_type": "code", "execution_count": 43, "id": "62538ac0-c6aa-4950-82e8-f53510f17d77", "metadata": {}, "outputs": [], "source": [ "response = client.diffprivlib.query(\n", " pipeline = dpl_pipeline,\n", " feature_columns = feature_columns,\n", " target_columns = target_columns,\n", " test_size = 0.1,\n", " test_train_split_seed = 1,\n", " dummy = False\n", ")" ] }, { "cell_type": "code", "execution_count": 44, "id": "7727b13a-cd2f-4f97-b550-88a4360fd601", "metadata": {}, "outputs": [], "source": [ "# Return the mean accuracy. \n", "model_score = response.result.score" ] }, { "cell_type": "code", "execution_count": 45, "id": "600d6c6e-7567-4564-92ad-d1538ac10af5", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "'The model has a mean accuracy of 0.41. It is a harsh metric because we are in a multi-label classification case.'" ] }, "execution_count": 45, "metadata": {}, "output_type": "execute_result" } ], "source": [ "f\"The model has a mean accuracy of {np.round(model_score, 2)}. It is a harsh metric because we are in a multi-label classification case.\"" ] }, { "cell_type": "code", "execution_count": 46, "id": "1824eea1-6ad8-4d2f-86d5-456a89318fef", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
Pipeline(steps=[('rf',\n",
       "                 RandomForestClassifier(accountant=BudgetAccountant(spent_budget=[(2.0, 0)]),\n",
       "                                        bounds=(array([  30.,   13.,  150., 2000.]),\n",
       "                                                array([  65.,   23.,  250., 7000.])),\n",
       "                                        classes=['Adelie', 'Chinstrap',\n",
       "                                                 'Gentoo'],\n",
       "                                        epsilon=2.0))])
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.
" ], "text/plain": [ "Pipeline(steps=[('rf',\n", " RandomForestClassifier(accountant=BudgetAccountant(spent_budget=[(2.0, 0)]),\n", " bounds=(array([ 30., 13., 150., 2000.]),\n", " array([ 65., 23., 250., 7000.])),\n", " classes=['Adelie', 'Chinstrap',\n", " 'Gentoo'],\n", " epsilon=2.0))])" ] }, "execution_count": 46, "metadata": {}, "output_type": "execute_result" } ], "source": [ "model = response.result.model\n", "model" ] }, { "cell_type": "code", "execution_count": 47, "id": "1e4a56bd-95bf-4355-81a6-eaf7d16f69c8", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "'For these feature values, the predicted species is is Adelie.'" ] }, "execution_count": 47, "metadata": {}, "output_type": "execute_result" } ], "source": [ "x_to_predict = pd.DataFrame({\n", " 'bill_length_mm': [30.0], 'bill_depth_mm': [20.0], 'flipper_length_mm': [170.0], 'body_mass_g': [5000.0]\n", "})\n", "predictions = model.predict(x_to_predict)[0]\n", "f\"For these feature values, the predicted species is is {predictions}.\"" ] }, { "cell_type": "markdown", "id": "09cb0eff-9864-42a3-a87a-cef2f7eca216", "metadata": {}, "source": [ "### d. Get a Synthetic Dataset with Smartnoise-Synth" ] }, { "cell_type": "markdown", "id": "790fe272-ae04-4daf-933e-6a9d86ac691f", "metadata": {}, "source": [ "Finally she gets a synthetic dataset to do the rest of her analysis. She chooses to only train on a subset on 3 columns: \"island\", \"bill_length_mm\" and \"bill_depth_mm\" but if we wanted she could train on the whole dataset.\n", "She also decides to use the `patectgan` synthesizer and keep all other default parameters." ] }, { "cell_type": "markdown", "id": "ded384a2-dcf6-43af-98b0-ae058b2a03ba", "metadata": {}, "source": [ "#### Train patectgan synthesizer on dummy data with Smartnoise-Synth" ] }, { "cell_type": "code", "execution_count": 48, "id": "ed292cec-8497-4e5b-b3bf-cca9227abf7d", "metadata": {}, "outputs": [], "source": [ "res_dummy = client.smartnoise_synth.query(\n", " synth_name=\"patectgan\",\n", " select_cols = [\"island\", \"bill_length_mm\", \"bill_depth_mm\"],\n", " epsilon=1.0,\n", " dummy=True,\n", ")\n", "dummy_synth_df = res_dummy.result.df_samples" ] }, { "cell_type": "code", "execution_count": 49, "id": "6fab49d9-f5c5-437e-b70f-234b84c2f7e5", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
islandbill_length_mmbill_depth_mm
0Dream41.06149717.365594
1Dream40.91610217.589716
2Dream55.32543519.580978
3Dream48.57402519.169669
4Biscoe48.33402316.253219
\n", "
" ], "text/plain": [ " island bill_length_mm bill_depth_mm\n", "0 Dream 41.061497 17.365594\n", "1 Dream 40.916102 17.589716\n", "2 Dream 55.325435 19.580978\n", "3 Dream 48.574025 19.169669\n", "4 Biscoe 48.334023 16.253219" ] }, "execution_count": 49, "metadata": {}, "output_type": "execute_result" } ], "source": [ "dummy_synth_df.head()" ] }, { "cell_type": "markdown", "id": "f75ef712-3d58-4548-904e-6f1da72493f7", "metadata": {}, "source": [ "#### Estimate cost of training patectgan synthesizer with Smartnoise-Synth" ] }, { "cell_type": "code", "execution_count": 50, "id": "82ab4aee-b4af-4fbd-93cc-b8171f7f9a52", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "CostResponse(epsilon=1.0, delta=0.00015673368198174188)" ] }, "execution_count": 50, "metadata": {}, "output_type": "execute_result" } ], "source": [ "res_cost = client.smartnoise_synth.cost(\n", " synth_name=\"patectgan\",\n", " epsilon=1.0,\n", " select_cols = [\"island\", \"bill_length_mm\", \"bill_depth_mm\"],\n", ")\n", "res_cost" ] }, { "cell_type": "markdown", "id": "804e2119-1513-45b2-a989-f5566e6c7bf4", "metadata": {}, "source": [ "#### Train patectgan synthesizer on private data with Smartnoise-Synth" ] }, { "cell_type": "code", "execution_count": 51, "id": "fbc8b354-e4db-4472-957b-468e768eddc4", "metadata": {}, "outputs": [], "source": [ "res = client.smartnoise_synth.query(\n", " synth_name=\"patectgan\",\n", " select_cols = [\"island\", \"bill_length_mm\", \"bill_depth_mm\"],\n", " epsilon=1.0,\n", " dummy=False,\n", ")\n", "synth_df = res.result.df_samples" ] }, { "cell_type": "code", "execution_count": 52, "id": "550fa89d-9537-4daf-9f96-42fa71f242b9", "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n", "\n", "\n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", " \n", "
islandbill_length_mmbill_depth_mm
0Dream51.67044215.726217
1Dream47.50582917.223182
2Torgersen56.44630915.642720
3Biscoe37.44280419.550416
4Dream46.18143115.531504
\n", "
" ], "text/plain": [ " island bill_length_mm bill_depth_mm\n", "0 Dream 51.670442 15.726217\n", "1 Dream 47.505829 17.223182\n", "2 Torgersen 56.446309 15.642720\n", "3 Biscoe 37.442804 19.550416\n", "4 Dream 46.181431 15.531504" ] }, "execution_count": 52, "metadata": {}, "output_type": "execute_result" } ], "source": [ "synth_df.head()" ] }, { "cell_type": "markdown", "id": "b3ae2259-3d83-46c6-a59c-9df178d98dc5", "metadata": {}, "source": [ "Out of curiosity, she checks the average bill length and variance of bill length on this dataset." ] }, { "cell_type": "code", "execution_count": 53, "id": "7d5a336e-80f0-48fa-84a3-33d0e51a2d3b", "metadata": {}, "outputs": [], "source": [ "synth_mean = np.round(synth_df[\"bill_length_mm\"].mean(), 2)\n", "synth_variance = np.round(synth_df[\"bill_length_mm\"].var(), 2)" ] }, { "cell_type": "code", "execution_count": 54, "id": "e890a8d9-0c7b-4805-be8a-81e66d5fa7ca", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "The average with Smartnoise-SQL on private data was 44.18.\n", "The average with Smartnoise-Synth on synthetic data is 44.34.\n" ] } ], "source": [ "print(\n", " f\"The average with Smartnoise-SQL on private data was {avg_bill_length}.\\n\"\n", " + f\"The average with Smartnoise-Synth on synthetic data is {synth_mean}.\"\n", ")" ] }, { "cell_type": "code", "execution_count": 55, "id": "11a14d9f-0fe3-4a5e-b425-d2192acd1e84", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "The variance with opendp on private data was 28.71.\n", "The variance with Smartnoise-Synth on synthetic data is 35.99.\n" ] } ], "source": [ "print(\n", " f\"The variance with opendp on private data was {var_bill_length}.\\n\"\n", " + f\"The variance with Smartnoise-Synth on synthetic data is {synth_variance}.\"\n", ")" ] }, { "cell_type": "markdown", "id": "94eaf59b-c108-424c-8978-b1c86e141ccb", "metadata": {}, "source": [ "## Step 4: See archives of queries" ] }, { "cell_type": "markdown", "id": "64003c53-de56-4bdc-a3c2-0c3e40031919", "metadata": {}, "source": [ "She now wants to verify all the queries that she did on the real data. It is possible because an archive of all queries is kept in a secure database. With a function call she can see her queries, budget and associated responses." ] }, { "cell_type": "code", "execution_count": 56, "id": "008fd230-cdfd-4e03-91ce-5a60b06c106d", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "4" ] }, "execution_count": 56, "metadata": {}, "output_type": "execute_result" } ], "source": [ "previous_queries = client.get_previous_queries()\n", "len(previous_queries)" ] }, { "cell_type": "code", "execution_count": 57, "id": "b712b269-64f2-4c7e-b8bf-d1a608933eff", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'user_name': 'Dr.Antartica',\n", " 'dataset_name': 'PENGUIN',\n", " 'dp_library': 'smartnoise_sql',\n", " 'client_input': {'dataset_name': 'PENGUIN',\n", " 'query_str': 'SELECT AVG(bill_length_mm) AS avg_bill_length_mm FROM df',\n", " 'epsilon': 0.5,\n", " 'delta': 0.0001,\n", " 'mechanisms': {},\n", " 'postprocess': True},\n", " 'response': {'epsilon': 1.0,\n", " 'delta': 4.999999999999449e-05,\n", " 'requested_by': 'Dr.Antartica',\n", " 'result': {'res_type': 'smartnoise_sql',\n", " 'df': {'index': [0],\n", " 'columns': ['avg_bill_length_mm'],\n", " 'data': [[44.182555797356954]],\n", " 'index_names': [None],\n", " 'column_names': [None]}}},\n", " 'timestamp': 1743433931.8376415}" ] }, "execution_count": 57, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# Smartnoise-SQL\n", "avg_bill_length_query = previous_queries[0]\n", "avg_bill_length_query" ] }, { "cell_type": "code", "execution_count": 58, "id": "8dfaf2b6-2b6c-480b-bcd7-250b0b2806a6", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'user_name': 'Dr.Antartica',\n", " 'dataset_name': 'PENGUIN',\n", " 'dp_library': 'opendp',\n", " 'client_input': {'dataset_name': 'PENGUIN',\n", " 'opendp_json': Measurement(\n", " input_domain = AtomDomain(T=String),\n", " input_metric = SymmetricDistance(),\n", " output_measure = MaxDivergence),\n", " 'fixed_delta': None,\n", " 'pipeline_type': 'legacy',\n", " 'mechanism': 'laplace'},\n", " 'response': {'epsilon': 0.7122093023265229,\n", " 'delta': 0.0,\n", " 'requested_by': 'Dr.Antartica',\n", " 'result': {'res_type': 'opendp', 'value': 28.71172413523004}},\n", " 'timestamp': 1743433951.408251}" ] }, "execution_count": 58, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# OpenDP\n", "var_bill_length_query = previous_queries[1]\n", "var_bill_length_query" ] }, { "cell_type": "code", "execution_count": 59, "id": "376315ec-6f38-4919-959e-d6bf244a4952", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'user_name': 'Dr.Antartica',\n", " 'dataset_name': 'PENGUIN',\n", " 'dp_library': 'diffprivlib',\n", " 'client_input': {'dataset_name': 'PENGUIN',\n", " 'diffprivlib_json': '{\"module\": \"diffprivlib\", \"version\": \"0.6.5\", \"pipeline\": [{\"type\": \"_dpl_type:RandomForestClassifier\", \"name\": \"rf\", \"params\": {\"n_estimators\": 10, \"n_jobs\": 1, \"random_state\": null, \"verbose\": 0, \"warm_start\": false, \"max_depth\": 5, \"epsilon\": 2.0, \"bounds\": {\"_tuple\": true, \"_items\": [[30.0, 13.0, 150.0, 2000.0], [65.0, 23.0, 250.0, 7000.0]]}, \"classes\": [\"Adelie\", \"Chinstrap\", \"Gentoo\"], \"shuffle\": false, \"accountant\": \"_dpl_instance:BudgetAccountant\"}}]}',\n", " 'feature_columns': ['bill_length_mm',\n", " 'bill_depth_mm',\n", " 'flipper_length_mm',\n", " 'body_mass_g'],\n", " 'target_columns': ['species'],\n", " 'test_size': 0.1,\n", " 'test_train_split_seed': 1,\n", " 'imputer_strategy': 'drop'},\n", " 'response': {'epsilon': 2.0,\n", " 'delta': 0.0,\n", " 'requested_by': 'Dr.Antartica',\n", " 'result': {'res_type': 'diffprivlib',\n", " 'score': 0.4117647058823529,\n", " 'model': Pipeline(steps=[('rf',\n", " RandomForestClassifier(accountant=BudgetAccountant(spent_budget=[(2.0, 0)]),\n", " bounds=(array([ 30., 13., 150., 2000.]),\n", " array([ 65., 23., 250., 7000.])),\n", " classes=['Adelie', 'Chinstrap',\n", " 'Gentoo'],\n", " epsilon=2.0))])}},\n", " 'timestamp': 1743433958.616408}" ] }, "execution_count": 59, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# DiffPrivLib\n", "reg_bill_length_query = previous_queries[2]\n", "reg_bill_length_query" ] }, { "cell_type": "code", "execution_count": 60, "id": "638817e0-d88f-407a-8136-210309651cf2", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'user_name': 'Dr.Antartica',\n", " 'dataset_name': 'PENGUIN',\n", " 'dp_library': 'smartnoise_synth',\n", " 'client_input': {'dataset_name': 'PENGUIN',\n", " 'synth_name': 'patectgan',\n", " 'epsilon': 1.0,\n", " 'delta': None,\n", " 'select_cols': ['island', 'bill_length_mm', 'bill_depth_mm'],\n", " 'synth_params': {},\n", " 'nullable': True,\n", " 'constraints': '',\n", " 'return_model': False,\n", " 'condition': '',\n", " 'nb_samples': 200},\n", " 'response': {'epsilon': 1.0,\n", " 'delta': 0.00015673368198174188,\n", " 'requested_by': 'Dr.Antartica',\n", " 'result': res_type \\\n", " index sn_synth_samples \n", " columns sn_synth_samples \n", " data sn_synth_samples \n", " index_names sn_synth_samples \n", " column_names sn_synth_samples \n", " \n", " df_samples \n", " index [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,... \n", " columns [island, bill_length_mm, bill_depth_mm] \n", " data [[Dream, 51.67044159024954, 15.726217210292816... \n", " index_names [None] \n", " column_names [None] },\n", " 'timestamp': 1743433982.805871}" ] }, "execution_count": 60, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# Smartnoise-Synth\n", "sysynth_query = previous_queries[3]\n", "sysynth_query" ] } ], "metadata": { "kernelspec": { "display_name": "venv", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.12.9" } }, "nbformat": 4, "nbformat_minor": 5 }