(#3) Defense: Add Cifar10 with Noise and Sql processing

src/utils/dataset.py

@@ -10,6 +10,7 @@
 from typings.dataset import NumpyDataset
 
 import tensorflow as tf
+import numpy as np
 
 keras = tf.keras
 
@@ -64,6 +65,26 @@ def postprocess(
         return (x_train, y_train), (x_test, y_test)
 
 
+class NoisySlqCifar10(ImageDataset, SlqMixin):
+    def __init__(self):
+        super().__init__(keras.datasets.cifar10)
+
+    def postprocess(
+        self, train: NumpyDataset, test: NumpyDataset
+    ) -> Tuple[NumpyDataset, NumpyDataset]:
+        def noisy(ds: np.array):
+            noise = 0.1 * np.random.normal(size=np.shape(ds))
+            return np.clip(ds + noise, 0.0, 1.0)
+
+        (x_train, y_train), (x_test, y_test) = train, test
+
+        (x_train, y_train), (x_test, y_test) = SlqMixin.process(
+            (noisy(x_train), y_train), (noisy(x_test), y_test)
+        )
+
+        return (x_train, y_train), (x_test, y_test)
+
+
 class IdemCifar10(ImageDataset, IdemMixin):
     def __init__(self):
         super().__init__(keras.datasets.cifar10)