OpenDP v0.5.0 (#486)

* Update to OpenDP v0.5

* approx bounds

* Pandas breaking change

Co-authored-by: Joshua <[email protected]>

Author: joshua-oss

.gitignore

@@ -112,6 +112,7 @@ poetry.lock
 # mlflow
 mlruns
 result.json
+run.py
 
 # VS Code
 .vscode/
@@ -135,4 +136,6 @@ PUMS_pid.csv
 PUMS_large.csv
 PUMS_dup.csv
 PUMS_null.csv
-*.db
+*.db
+
+

sql/HISTORY.md

@@ -1,3 +1,10 @@
+# SmartNoise SQL v0.2.5 Release Notes
+
+* Update to use OpenDP v0.5, support for Mac Silicon
+* Switch to use discrete Laplace for all integer sums and counts
+* Enable discrete Gaussian option
+* `get_privacy_cost` now allows lists of queries
+
 # SmartNoise SQL v0.2.4 Release Notes
 
 * Support for BigQuery (thanks, @oskarkocol!)

sql/README.md

@@ -138,23 +138,6 @@ epsilon_many, _ = reader.odometer.spent
 print(f'{epsilon_many} < {epsilon_single * 100}')
 ```
 
-## Accuracy
-
-The `get_simple_accuracy` method returns the column-wise accuracies for a given alpha for a given query.
-
-```python
-privacy = Privacy(epsilon=1.0, delta=10e-6)
-
-reader = from_connection(conn, metadata=metadata, privacy=privacy)
-
-query = 'SELECT COUNT(*) AS n, SUM(age) AS age FROM PUMS.PUMS'
-
-acc95 = reader.get_simple_accuracy(query, alpha=0.05)
-print(f'n will be +/- {acc95[0]} in 95% of executions.  Age will be +/- {acc95[1]}')
-```
-
-This method only returns simple accuracies, where the noise scale for each column is fixed and does not vary per row.  Statistics like AVG and VARIANCE are computed from a quotient of noisy sum and noisy count, so the accuracy can vary widely per row.  In these cases, a per-row accuracy can be obtained with `execute_with_accuracy`.  
-
 ## Histograms
 
 SQL `group by` queries represent histograms binned by grouping key.  Queries over a grouping key with unbounded or non-public dimensions expose privacy risk. For example:

sql/VERSION

@@ -1 +1 @@
-0.2.4
+0.2.5

sql/docs/source/advanced.rst

@@ -16,12 +16,10 @@ You can override the default mechanisms used for differentially private summary
     privacy = Privacy(epsilon=1.0)
     print(f"We default to using {privacy.mechanisms.map[Stat.count]} for counts.")
     print("Switching to use gaussian")
-    privacy.mechanisms.map[Stat.count] = Mechanism.gaussian
+    privacy.mechanisms.map[Stat.count] = Mechanism.discrete_gaussian
 
 The list of statistics that can be mapped is in the ``Stat`` enumeration, and the mechanisms available are listed in the ``Mechanism`` enumeration.  The ``AVG`` sumamry statistic is computed from a sum and a count, each of which can be overriden.
 
-For integer sums, you can specify ``Stat.sum_int``, and specify ``Stat.sum_large_int`` separately for large integer sums.  A "large" integer sum is any sum that is greater than the value set in ``mechanisms.large``, which defaults to 1000.  This is primarily useful when using geometric mechanism for integer sums, since the geometric mechanism slows down for large integer sums.
-
 pre_aggregated
 --------------
 

sql/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "smartnoise-sql"
-version = "0.2.4"
+version = "0.2.5"
 description = "Differentially Private SQL Queries"
 authors = ["SmartNoise Team <[email protected]>"]
 license = "MIT"
@@ -11,7 +11,7 @@ readme = "README.md"
 
 [tool.poetry.dependencies]
 python = ">3.6,<3.11"
-opendp = "^0.4.0"
+opendp = "^0.5.0"
 antlr4-python3-runtime = "4.9.3"
 pandasql = "^0.7.3"
 PyYAML = "^5.4.1"

sql/setup.py

@@ -1,7 +1,6 @@
-from .gaussian import Gaussian
 from .laplace import Laplace
-from .geometric import Geometric
-from .analytic_gaussian import AnalyticGaussian
+from .discrete_laplace import DiscreteLaplace
+from .discrete_gaussian import DiscreteGaussian
 from .base import Mechanism, Unbounded
 
-__all__ = ["Gaussian", "Laplace", "Geometric", "Mechanism", "Unbounded", "AnalyticGaussian"]
+__all__ = ["Laplace", "DiscreteLaplace", "DiscreteGaussian", "Mechanism", "Unbounded",]