Upload 7 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+car_on_road.png filter=lfs diff=lfs merge=lfs -text

app.py

@@ -0,0 +1,95 @@
+import torch
+import torchvision
+import torchvision.transforms as transforms
+import torch.nn.functional as F
+from PIL import Image
+import streamlit as st
+from pathlib import Path
+import requests
+
+# Load pre-trained ResNet model
+model = torchvision.models.resnet50(pretrained=True)
+model.eval()
+
+# Load ImageNet class labels
+LABELS_URL = "https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt"
+labels = requests.get(LABELS_URL).text.strip().split("\n")
+
+# Image transform
+transform = transforms.Compose([
+    transforms.Resize((224, 224)),
+    transforms.ToTensor(),
+])
+
+# Load predefined signs from 'signs' folder
+signs_dir = Path("signs")
+presets = {f.stem.replace("_", " "): f for f in signs_dir.glob("*.")}
+
+def preprocess_image(image):
+    image = image.convert("RGB")
+    tensor = transform(image).unsqueeze(0)
+    return tensor
+
+def predict_class(tensor):
+    with torch.no_grad():
+        outputs = model(tensor)
+        _, predicted = outputs.max(1)
+        return labels[predicted.item()]
+
+def generate_adversarial_example(tensor, epsilon, target_class=400):
+    tensor.requires_grad = True
+    outputs = model(tensor)
+    loss = F.cross_entropy(outputs, torch.tensor([target_class]))
+    loss.backward()
+    perturbation = epsilon * tensor.grad.sign()
+    perturbed_tensor = torch.clamp(tensor + perturbation, 0, 1)
+    return perturbed_tensor
+
+def main():
+    st.title("Adversarial Attack on Traffic Signs")
+    st.write("Upload a traffic sign image or select a predefined one, then apply perturbation.")
+
+    col1, col2 = st.columns(2)
+
+    with col1:
+        uploaded_file = st.file_uploader("Upload a traffic sign image", type=["png", "jpg", "jpeg"])
+    with col2:
+        selected_preset = st.selectbox("Or choose a predefined sign", [None] + list(presets.keys()))
+
+    # User-defined perturbation strength limit
+    max_epsilon = st.number_input("Set Maximum Perturbation Strength (ε)", min_value=0.01, max_value=1.0, value=0.1, step=0.01)
+    epsilon = st.slider("Select Perturbation Strength (ε)", 0.0, float(max_epsilon), 0.01, step=0.01)
+
+    def load_image():
+        if uploaded_file:
+            return Image.open(uploaded_file)
+        elif selected_preset:
+            return Image.open(presets[selected_preset])
+        else:
+            return None
+
+    image = load_image()
+
+    if image:
+        st.image(image, caption="Original Image", use_container_width=True)
+        tensor = preprocess_image(image)
+        original_label = predict_class(tensor)
+
+        perturbed_tensor = generate_adversarial_example(tensor.clone(), epsilon)
+        perturbed_image = transforms.ToPILImage()(perturbed_tensor.squeeze())
+        adversarial_label = predict_class(perturbed_tensor)
+
+        col1, col2 = st.columns(2)
+        with col1:
+            st.markdown("### Original Prediction")
+            st.success(original_label)
+        with col2:
+            st.markdown("### Adversarial Prediction")
+            st.error(adversarial_label)
+
+        st.image(perturbed_image, caption="Perturbed Image", use_container_width=True)
+    else:
+        st.warning("Please upload image.")
+
+if __name__ == "__main__":
+    main()

requrements.txt

@@ -0,0 +1,5 @@
+torch
+torchvision
+Pillow
+streamlit
+requests