I am planning to make an app that serves ads using TFLite and CoreML so the user privacy could be protected. Is there any limitations to this?
Perhaps I am missing something. How would you train the ML model without a dataset of aggregated user data?
Generally, you will have two implimentations of this:
- The app passes raw (hopefully secure) data to a server to handle both training and predicting. If the the model is being further trained with the extra information (Kite is an excellent example of this), then it would not be beneficial to have the model on the users machine, because they could not make use of other users’ data.
- The app passes data to a server to handle predicting.
The limitation of apps that train and/or process, is resource intensiveness. I cannot imagine this would be efficient on a phone.
I thought of using pre-trained models and retraining them based on my custom data using transfer learning.
Can’t models that are bundled in apps be updated with user data on-device, helping models stay relevant to user behavior without compromising privacy?
BTW I’m a beginner and slowly learning things
From what I understand, no. Not without the same resource use as needed in the cloud. As I eluded at in my previous response, it is a toss-up between an app that is resource intensive, and one that is server dependant.
Keep up with your venture. You will learn a lot, and you could produce a useful app. Who knows what you will find.
that’s a good idea. Anyway Internet privacy is a serious issue thse days. That’s why I started using VeePN to protect my data. It works well and helps me to bypass all restrictions.