Add geometric updating for quantile-based adaptive clipping.

PiperOrigin-RevId: 285799976

tensorflow_privacy/privacy/dp_query/BUILD

@@ -127,6 +127,7 @@ py_test(
     deps = [
         ":quantile_adaptive_clip_sum_query",
         ":test_utils",
+        "//third_party/py/absl/testing:parameterized",
         "//third_party/py/numpy",
         "//third_party/py/tensorflow",
         "//third_party/py/tensorflow_privacy/privacy/analysis:privacy_ledger",

tensorflow_privacy/privacy/dp_query/quantile_adaptive_clip_sum_query.py

@@ -66,7 +66,8 @@ class QuantileAdaptiveClipSumQuery(dp_query.DPQuery):
       target_unclipped_quantile,
       learning_rate,
       clipped_count_stddev,
-      expected_num_records):
+      expected_num_records,
+      geometric_update=False):
     """Initializes the QuantileAdaptiveClipSumQuery.
 
     Args:
@@ -84,6 +85,7 @@ class QuantileAdaptiveClipSumQuery(dp_query.DPQuery):
         should be about 0.5 for reasonable privacy.
       expected_num_records: The expected number of records per round, used to
         estimate the clipped count quantile.
+      geometric_update: If True, use geometric updating of clip.
     """
     self._initial_l2_norm_clip = initial_l2_norm_clip
     self._noise_multiplier = noise_multiplier
@@ -107,6 +109,8 @@ class QuantileAdaptiveClipSumQuery(dp_query.DPQuery):
         sum_stddev=clipped_count_stddev,
         denominator=expected_num_records)
 
+    self._geometric_update = geometric_update
+
   def set_ledger(self, ledger):
     """See base class."""
     self._sum_query.set_ledger(ledger)
@@ -214,8 +218,12 @@ class QuantileAdaptiveClipSumQuery(dp_query.DPQuery):
     # the true quantile matches the target.
     loss_grad = unclipped_quantile - global_state.target_unclipped_quantile
 
-    new_l2_norm_clip = gs.l2_norm_clip - global_state.learning_rate * loss_grad
+    update = global_state.learning_rate * loss_grad
-    new_l2_norm_clip = tf.maximum(0.0, new_l2_norm_clip)
+
+    if self._geometric_update:
+      new_l2_norm_clip = gs.l2_norm_clip * tf.math.exp(-update)
+    else:
+      new_l2_norm_clip = tf.math.maximum(0.0, gs.l2_norm_clip - update)
 
     new_sum_stddev = new_l2_norm_clip * global_state.noise_multiplier
     new_sum_query_global_state = self._sum_query.make_global_state(

tensorflow_privacy/privacy/dp_query/quantile_adaptive_clip_sum_query_test.py

@@ -18,6 +18,8 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+from absl.testing import parameterized
+
 import numpy as np
 import tensorflow as tf
 
@@ -28,7 +30,8 @@ from tensorflow_privacy.privacy.dp_query import test_utils
 tf.compat.v1.enable_eager_execution()
 
 
-class QuantileAdaptiveClipSumQueryTest(tf.test.TestCase):
+class QuantileAdaptiveClipSumQueryTest(
+    tf.test.TestCase, parameterized.TestCase):
 
   def test_sum_no_clip_no_noise(self):
     record1 = tf.constant([2.0, 0.0])
@@ -158,6 +161,42 @@ class QuantileAdaptiveClipSumQueryTest(tf.test.TestCase):
       self.assertAllClose(actual_clip.numpy(), expected_clip)
       self.assertAllClose(actual_sum.numpy(), (expected_sum,))
 
+  def test_adaptation_target_zero_geometric(self):
+    record1 = tf.constant([5.0])
+    record2 = tf.constant([-2.5])
+
+    query = quantile_adaptive_clip_sum_query.QuantileAdaptiveClipSumQuery(
+        initial_l2_norm_clip=16.0,
+        noise_multiplier=0.0,
+        target_unclipped_quantile=0.0,
+        learning_rate=np.log(2.0),      # Geometric steps in powers of 2.
+        clipped_count_stddev=0.0,
+        expected_num_records=2.0,
+        geometric_update=True)
+
+    global_state = query.initial_global_state()
+
+    initial_clip = global_state.l2_norm_clip
+    self.assertAllClose(initial_clip, 16.0)
+
+    # For two iterations, nothing is clipped, so the clip is cut in half.
+    # Then one record is clipped, so the clip goes down by only sqrt(2.0) to
+    # 4 / sqrt(2.0). Still only one record is clipped, so it reduces to 2.0.
+    # Now both records are clipped, and the clip norm stays there (at 2.0).
+
+    four_div_root_two = 4 / np.sqrt(2.0)   # approx 2.828
+
+    expected_sums = [2.5, 2.5, 1.5, four_div_root_two - 2.5, 0.0]
+    expected_clips = [8.0, 4.0, four_div_root_two, 2.0, 2.0]
+    for expected_sum, expected_clip in zip(expected_sums, expected_clips):
+      actual_sum, global_state = test_utils.run_query(
+          query, [record1, record2], global_state)
+
+      actual_clip = global_state.l2_norm_clip
+
+      self.assertAllClose(actual_clip.numpy(), expected_clip)
+      self.assertAllClose(actual_sum.numpy(), (expected_sum,))
+
   def test_adaptation_target_one(self):
     record1 = tf.constant([-1.5])
     record2 = tf.constant([2.75])
@@ -191,22 +230,64 @@ class QuantileAdaptiveClipSumQueryTest(tf.test.TestCase):
       self.assertAllClose(actual_clip.numpy(), expected_clip)
       self.assertAllClose(actual_sum.numpy(), (expected_sum,))
 
-  def test_adaptation_linspace(self):
+  def test_adaptation_target_one_geometric(self):
+    record1 = tf.constant([-1.5])
+    record2 = tf.constant([3.0])
+
+    query = quantile_adaptive_clip_sum_query.QuantileAdaptiveClipSumQuery(
+        initial_l2_norm_clip=0.5,
+        noise_multiplier=0.0,
+        target_unclipped_quantile=1.0,
+        learning_rate=np.log(2.0),      # Geometric steps in powers of 2.
+        clipped_count_stddev=0.0,
+        expected_num_records=2.0,
+        geometric_update=True)
+
+    global_state = query.initial_global_state()
+
+    initial_clip = global_state.l2_norm_clip
+    self.assertAllClose(initial_clip, 0.5)
+
+    # On the first two iterations, both are clipped, so the clip is doubled.
+    # When the clip reaches 2.0, only one record is clipped, so the clip is
+    # multiplied by sqrt(2.0). Still only one is clipped so it increases to 4.0.
+    # Now both records are clipped, and the clip norm stays there (at 4.0).
+
+    two_times_root_two = 2 * np.sqrt(2.0)   # approx 2.828
+
+    expected_sums = [0.0, 0.0, 0.5, two_times_root_two - 1.5, 1.5]
+    expected_clips = [1.0, 2.0, two_times_root_two, 4.0, 4.0]
+    for expected_sum, expected_clip in zip(expected_sums, expected_clips):
+      actual_sum, global_state = test_utils.run_query(
+          query, [record1, record2], global_state)
+
+      actual_clip = global_state.l2_norm_clip
+
+      self.assertAllClose(actual_clip.numpy(), expected_clip)
+      self.assertAllClose(actual_sum.numpy(), (expected_sum,))
+
+  @parameterized.named_parameters(
+      ('start_low_arithmetic', True, False),
+      ('start_low_geometric', True, True),
+      ('start_high_arithmetic', False, False),
+      ('start_high_geometric', False, True))
+  def test_adaptation_linspace(self, start_low, geometric):
     # 100 records equally spaced from 0 to 10 in 0.1 increments.
     # Test that with a decaying learning rate we converge to the correct
-    # median with error at most 0.1.
+    # median value and bounce around it.
     records = [tf.constant(x) for x in np.linspace(
         0.0, 10.0, num=21, dtype=np.float32)]
 
     learning_rate = tf.Variable(1.0)
 
     query = quantile_adaptive_clip_sum_query.QuantileAdaptiveClipSumQuery(
-        initial_l2_norm_clip=0.0,
+        initial_l2_norm_clip=(1.0 if start_low else 10.0),
         noise_multiplier=0.0,
         target_unclipped_quantile=0.5,
         learning_rate=learning_rate,
         clipped_count_stddev=0.0,
-        expected_num_records=2.0)
+        expected_num_records=2.0,
+        geometric_update=geometric)
 
     global_state = query.initial_global_state()
 
@@ -219,20 +300,24 @@ class QuantileAdaptiveClipSumQueryTest(tf.test.TestCase):
       if t > 40:
         self.assertNear(actual_clip, 5.0, 0.25)
 
-  def test_adaptation_all_equal(self):
+  @parameterized.named_parameters(
-    # 100 equal records. Test that with a decaying learning rate we converge to
+      ('arithmetic', False),
+      ('geometric', True))
+  def test_adaptation_all_equal(self, geometric):
+    # 20 equal records. Test that with a decaying learning rate we converge to
     # that record and bounce around it.
     records = [tf.constant(5.0)] * 20
 
     learning_rate = tf.Variable(1.0)
 
     query = quantile_adaptive_clip_sum_query.QuantileAdaptiveClipSumQuery(
-        initial_l2_norm_clip=0.0,
+        initial_l2_norm_clip=1.0,
         noise_multiplier=0.0,
         target_unclipped_quantile=0.5,
         learning_rate=learning_rate,
         clipped_count_stddev=0.0,
-        expected_num_records=2.0)
+        expected_num_records=2.0,
+        geometric_update=geometric)
 
     global_state = query.initial_global_state()
 
@@ -243,7 +328,7 @@ class QuantileAdaptiveClipSumQueryTest(tf.test.TestCase):
       actual_clip = global_state.l2_norm_clip
 
       if t > 40:
-        self.assertNear(actual_clip, 5.0, 0.25)
+        self.assertNear(actual_clip, 5.0, 0.5)
 
   def test_ledger(self):
     record1 = tf.constant([8.5])