Update app.py

app.py

@@ -9,6 +9,7 @@ from PIL import Image
 
 # I/O image dimensions for display
 DIMS = (100, 100)
+
 # Load the trained model
 model = load_model('mnist_model.h5')
 
@@ -78,9 +79,39 @@ def gradio_mask(image, steps, increment=1):
     modified_image, original_label, predicted_label = progressively_mask_image(image, steps, increment)
     return modified_image, f"Original Label: {original_label}, New Label: {predicted_label}"
 
+# FGSM attack function
+def fgsm_attack(image, epsilon, data_grad):
+    sign_of_grad = tf.sign(data_grad)
+    perturbed_image = image + epsilon * sign_of_grad
+    perturbed_image = tf.clip_by_value(perturbed_image, 0, 1)
+    return perturbed_image
+
+# Create adversarial example function
+def create_adversarial_pattern(input_image, input_label):
+    with tf.GradientTape() as tape:
+        tape.watch(input_image)
+        prediction = model(input_image)
+        loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)(input_label, prediction)
+    gradient = tape.gradient(loss, input_image)
+    return gradient
+
+# Generate adversarial examples
+epsilon = 0.1  # Tweak epsilon to change the intensity of perturbation
+adversarial_examples = []
+
+for i in range(10):
+    img = tf.convert_to_tensor(test_images[i:i+1])
+    label = test_labels[i:i+1]
+    label = tf.reshape(label, [1, 1])
+
+    perturbations = create_adversarial_pattern(img, label)
+    adv_x = fgsm_attack(img, epsilon, perturbations)
+    adversarial_examples.append([cv2.resize(adv_x.numpy().squeeze(), DIMS)])
+
 class GradioInterface:
     def __init__(self):
         self.preloaded_examples = mnist_examples
+        self.adversarial_examples = adversarial_examples
 
     def create_interface(self):
         app_styles = """
@@ -224,12 +255,16 @@ class GradioInterface:
                 with gr.Column():
                     input_image = gr.Image(label="Input Image", type="pil", format="png", elem_classes="image-preview")
                     steps_input = gr.Slider(minimum=1, maximum=100, label="Attributions Drop Percentage", step=1, value=5)
-                    #increment_input = gr.Slider(minimum=1, maximum=100, label="increment", step=1, value=5)
                     examples = gr.Examples(
                         examples=self.preloaded_examples,
                         inputs=input_image,
                         label="MNIST Examples"
                     )
+                    adv_examples = gr.Examples(
+                        examples=self.adversarial_examples,
+                        inputs=input_image,
+                        label="Adversarial Examples"
+                    )
                 with gr.Column():
                     result = gr.Image(label="Result", elem_classes="image-preview")
                     prediction_details = gr.Textbox(label="Prediction Details")
@@ -237,7 +272,6 @@ class GradioInterface:
 
             run_button.click(
                 fn=gradio_mask,
-                #inputs=[input_image, steps_input, increment_input],
                 inputs=[input_image, steps_input],
                 outputs=[result, prediction_details],
             )