The Manipulated Image Detector is a free Chrome extension developed by Nicholas Gin. It is designed to predict whether an image on a webpage is authentic or if it has been manipulated. Utilizing an on-device model built with TensorFlow.js scripts, this extension offers a convenient way to assess the authenticity of images right from your Chrome browser.

To use the Manipulated Image Detector, simply right-click on the desired image on a webpage and select "Predict Image Authenticity" from the Context Menu. A pop-up window will appear, displaying the detector. The image will be quickly analyzed, and the detector will provide a prediction regarding its authenticity. The window will remain active and update in real-time whenever a new image is uploaded to the detector.

The model used by the Manipulated Image Detector is a custom convolutional neural network trained on a dataset of 3,938 authentic images and 3,938 manipulated images from the CASIA2 dataset. It leverages the distinguishing variables learned during the training process to make predictions. However, it's important to note that the model is not foolproof and achieved an F1 score of approximately 87.3% when tested on a separate subset of images from the CASIA2 dataset. Therefore, while the predictions made by the model are informed, they should be used as a tool rather than taken as absolute fact.