Lovable Prompt Creator - AI Prompt Generator

## APPLICATION OVERVIEW
The application is a web-based interactive dashboard designed to analyze the adversarial robustness of vision models, specifically implementing the Rectangular Occlusion Attack (ROA) framework. Its primary purpose is to provide users with insightful metrics and visualizations that demonstrate how various models respond to adversarial attacks.

## CORE FEATURES
1. **Model Selection**: A dropdown menu for users to choose from pre-trained models like ResNet-50, VGG-16, and ViT.
2. **Dataset Configuration**: A file uploader that enables users to upload target images (size constraints of 224x224 or 32x32).
3. **Attack Parameters**: Interactive sliders for adjusting sticker dimensions, a radio button for search strategies, and input fields for intensity and iterations.
4. **Metric Scorecards**: Prominent displays for key performance indicators such as Adversarial Accuracy, Attack Success Rate (ASR), and Empirical Robustness.
5. **Visual Comparison Module**: A side-by-side display of the original image, the adversarial mask, and the attacked image, showcasing predicted labels and confidence scores.
6. **Strategic Mitigation Report**: An expandable section that suggests possible defenses based on the analysis results.

## DESIGN SPECIFICATIONS
- **Visual Style**: Minimalist - The design focuses on a clean, simple aesthetic with ample white space, ensuring that the most important information stands out.
- **Color Mode**: Light theme with dark text on light backgrounds, utilizing consistent color coding (green for 'Robust', red for 'Vulnerable').
- **Layout**: A clear sidebar on the left for global configurations and a main dashboard area for displaying metrics and visual results. High-impact KPIs are positioned at the top for immediate visibility.
- **Typography**: Use sans-serif fonts for clarity and modernity. Headings should be bold for emphasis, while body text remains lightweight for easy reading.

## TECHNICAL REQUIREMENTS
- **Framework**: Python-based Streamlit for rapid web app development.
- **UI Components**: Utilize Streamlit components for inputs and visualizations.
- **Data Handling**: Implement asynchronous data loading with a loading spinner to manage heavy gradient computations effectively.

## IMPLEMENTATION STEPS
1. **Set up the Streamlit environment**: Install Streamlit and necessary libraries (torch, torchvision, adversarial-robustness-toolbox).
2. **Create Sidebar Components**: 
   - Implement a dropdown for model selection.
   - Add a file uploader for target images.
   - Integrate sliders and radio buttons for attack parameters.
3. **Main Dashboard Layout**: 
   - Structure the main area to display metric scorecards at the top.
   - Arrange a visual comparison module using `st.columns` to show images side-by-side.
4. **Implement Metric Calculations**: Calculate adversarial accuracy, ASR, and robustness metrics based on model predictions and comparisons.
5. **Visualizations**: 
   - Use Plotly to create interactive graphics for the 'Security Evaluation Curve'.
   - Implement a saliency heatmap using Captum or Eigen-CAM.
6. **Dynamic Mitigation Report**: 
   - Create an expandable section that suggests defenses based on user input and results dynamically.
7. **Asynchronous Data Handling**: Ensure that the application handles data loading efficiently, displaying a loading spinner during heavy computations.

## USER EXPERIENCE
Users will interact with the sidebar to configure their model and dataset, adjusting attack parameters as necessary. Upon uploading an image and initiating the analysis, they will see loading indicators followed by a comprehensive dashboard displaying key metrics and visual comparisons. The minimalist design ensures that users can focus on the data and insights without distractions, while the dynamic mitigation report provides actionable suggestions based on their findings.