Merge pull request #53 from M-N-Tran/dev

Implement the spatial smoothing defence

Author: MARIA NICOLAE

src/defences/spatial_smoothing.py

@@ -0,0 +1,63 @@
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+from scipy import ndimage
+
+from src.defences.preprocessor import Preprocessor
+
+
+class SpatialSmoothing(Preprocessor):
+    """
+    Implement the local spatial smoothing defence approach.
+    Defence method from https://arxiv.org/abs/1704.01155.
+    """
+    params = ["window_size"]
+
+    def __init__(self, window_size=3):
+        """
+        Create an instance of local spatial smoothing.
+        :param window_size: (int) The size of the sliding window.
+        """
+        self.is_fitted = True
+        self.set_params(window_size=window_size)
+
+    def __call__(self, x_val, window_size=3):
+        """
+        Apply local spatial smoothing to sample x_val.
+        :param x_val: (np.ndarray) Sample to smooth. `x_val` is supposed to
+        have shape (batch_size, width, height, depth).
+        :param window_size: (int) The size of the sliding window.
+        :return: Smoothed sample
+        :rtype: np.ndarray
+        """
+        self.set_params(window_size=window_size)
+        size = (1, window_size, window_size, 1)
+        result = ndimage.filters.median_filter(x_val, size=size, mode="reflect")
+
+        return result
+
+    def fit(self, x_val, y_val=None, **kwargs):
+        """
+        No parameters to learn for this method; do nothing.
+        """
+        pass
+
+    def set_params(self, **kwargs):
+        """
+        Take in a dictionary of parameters and applies defense-specific checks
+        before saving them as attributes.
+        Defense-specific parameters:
+        :param window_size: (int) The size of the sliding window.
+        """
+        # Save attack-specific parameters
+        super(SpatialSmoothing, self).set_params(**kwargs)
+
+        if type(self.window_size) is not int or self.window_size <= 0:
+            raise ValueError("Sliding window size must be a positive integer")
+
+        return True
+
+
+
+
+
+

src/defences/spatial_smoothing_unittest.py

@@ -0,0 +1,44 @@
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import unittest
+
+import numpy as np
+
+from src.defences.spatial_smoothing import SpatialSmoothing
+
+
+class TestLocalSpatialSmoothing(unittest.TestCase):
+    def test_ones(self):
+        m, n = 10, 2
+        x = np.ones((1, m, n, 3))
+
+        # Start to test
+        for window_size in range(1, 20):
+            with self.subTest("Sliding window size = {}".format(window_size)):
+                preprocess = SpatialSmoothing()
+                smoothed_x = preprocess(x, window_size)
+                self.assertTrue((smoothed_x == 1).all())
+
+    def test_fix(self):
+        x = np.array([[[[1], [2], [3]], [[7], [8], [9]], [[4], [5], [6]]]])
+
+        # Start to test
+        preprocess = SpatialSmoothing()
+        smoothed_x = preprocess(x, 3)
+        self.assertTrue((smoothed_x==np.array(
+            [[[[2], [3], [3]], [[4], [5], [6]], [[5], [6], [6]]]])).all())
+
+        smoothed_x = preprocess(x, 1)
+        self.assertTrue((smoothed_x==x).all())
+
+        smoothed_x = preprocess(x, 2)
+        self.assertTrue((smoothed_x==np.array(
+            [[[[1], [2], [3]], [[7], [7], [8]], [[7], [7], [8]]]])).all())
+
+
+if __name__ == '__main__':
+    unittest.main()
+
+
+
+