Inspiration
With the rapid advancement of generative AI tools, AI-generated images are becoming increasingly realistic and difficult to distinguish from real photographs. This creates risks in misinformation, digital fraud, and deepfake manipulation. Our goal was to build a practical solution that helps users instantly verify whether an image is authentic or AI-generated.
What it does
The application allows users to upload an image through a simple web interface. The system analyzes the image using a trained deep learning model and predicts whether it is: • Real Image • AI-Generated Image The result is displayed instantly with a prediction confidence.
How we built it
1.Collected datasets of real and AI-generated images 2.Preprocessed images (resizing to 224x224, normalization) 3.Used Transfer Learning with MobileNetV2 for feature extraction 4.Added a binary classification layer for prediction 5.Trained the model using TensorFlow/Keras Saved the trained model (. h5 format) 6.Built a Streamlit web interface for user interaction 7.Deployed the application using streamlit Cloud
Challenges we ran into
Model initially produced random predictions due to small dataset • Class label mismatch caused reversed predictions • Limited training data led to overfitting • Deployment issues with model file size We resolved these by: • Using transfer learning • Increasing dataset size • Adjusting prediction logic • Optimizing deployment setup
Accomplishments that we're proud of
We are proud of building a fully functional AI image detector within the hackathon timeframe, integrating deep learning with a live web application, and successfully deploying it online to address real-world misinformation challenges.
What we learned
Importance of data preprocessing • Handling class imbalance in ML models • Deploying ML models as web applications • Practical challenges of detecting AI-generated content
What's next for Untitled
TruthLens AI