Methods and systems for learning noise distributions for neural networks

A method of transmitting data over a network, comprising:

• generating source data to be transmitted over the network;
• performing a conversion operation on the source data to convert the source data into converted data such that the source data is not recoverable from the converted data;
• and
• transmitting the converted data over the network,
• wherein the conversion operation includes:
• selecting, as a local network, one or more layers of a deep neural network (DNN) with multiple layers between input and output layers such that a computation of the one or more layers is made on an edge device configured to control data flow at a boundary between the local network and the DNN and selecting, as a remote network, remaining layers of the DNN;
• obtaining a first activation tensor by providing an input to the local network and obtaining an output of the local network responsive to the input;
• obtaining a second activation tensor by injecting a noise tensor sampled from stored distributions to the first activation tensor;
• feeding the first activation tensor and the second activation tensor to the remote network to obtain first and second results, respectively, from the remote network;
• finding a trained noise tensor that minimizes a loss function of the DNN;
• adding the trained noise tensor to a third activation to be transmitted to the remote network;
• and
• transmitting the third activation.