Privacy · Feyn

by: Kevin Broløs
(Feyn version 2.1 or newer)

Privacy is key to us here at Abzu. For that reason, we only send information to the QLattice that is needed for it to deliver better models to you, and only when you explicitly ask Feyn to (or use auto_run).

The information sent to the QLattice when you update therefore only consists of:

The token and QLattice identifiers when you connect.
The codes of your features (this is translated locally based on the order of your features in the dataset - the server doesn't know which code corresponds to what)
The model structures you want to reinforce, which is just a list of codes that denote the features and functions.
Where this structure originated in the QLattice, which is just to help us reinforce and evolve that specific evolution.

None of your data is at any point exchanged and does not leave your machine.

What exactly gets sent to the QLattice

Here's a concrete example of models trained using auto_run.

Each epoch of auto_run calls two functions under the hood that communicate with the QLattice:

QLattice.update, where it takes the best models, converts them to a JSON payload and sends it to the QLattice.
QLattice.sample_models, where you provide feature names and the type of model you want to sample. The feature names are translated to codes, so the QLattice doesn't care what you're building models about, only what relates and what works.

(notice that auto_run is a method on the QLattice object to denote that it has remote connection).

import feyn

train, test = feyn.datasets.make_classification()

ql = feyn.connect_qlattice()
best_models = ql.auto_run(train, 'y', kind='classification', max_complexity=3)

best_models[0]

In the manual workflow, it's more clear where the connection points with the QLattice happen:

    import feyn
    ql = feyn.connect_qlattice()

    # Connection happens here to sample models with the requested features
    models = ql.sample_models(train.columns, output_name='y', kind = 'classification', max_complexity=3)

    # This is all performed locally
    models = feyn.fit_models(models, train, criterion='bic')
    models = feyn.prune_models(models)

    # Connection happens here to update the QLattice
    ql.update(models)

The ql.update function takes each model in the list, converts it to json, and sends the resulting model structures to the QLattice in this format:

    {
        'codes': [2002, 10008, 10003, 1000, 10014, 10005, 10018, 2000, 1002, 10000, 1007, 2001, 2001, 10009, 10004, 2000, 2001, 1006, 1004, 2002, 2002, 2001, 10006, 2001, 1006, 10015, 2001, 2001, 2002, 10003, 10008, 10017],
        'data': {
            'pid': 5452227153675128204,
            'ppid': 15462473290879388848,
            'generation': 3,
            'action': 'd',
            'ix': 2
        },
        'qid': 11
    }