No personal data (images, metadata or AI learnings) ever leaves your local server if your run your own instance of Photonix.
We were able to develop face recognition without sending any data to the cloud and break it down so the only part that needs retraining specifically to your collection of faces is very fast. You can read the in-depth steps of our face detection and recognition model here if you are so inclined.
All the computationally intensive steps mentioned in that document are performed by pre-trained, industry-standard models. This gets us to the point where we have a fingerprint (or embedding) array of 128 numbers to roughly represent each detected face.
The very last step labelled Similarity Calculation / Clustering / Classification is the only part of the model that is built on the local machine. This similarity index is very quick to build (typically a few seconds) and just means that we don't have to compare the current face embedding individually with every other face embedding already in the system when we want to find the closest match.