Learning from distributed data

A computer system for training a classifier on confidential data in a first data set and a second data set, the computer system comprising:

• one or more processors;
• a first data collection device configured to:
• store the confidential data in the first data set comprising multiple first data samples, each of the multiple first data samples comprising multiple first feature values and a first label that classifies that first data sample;
• create first non-confidential training data by determining each of multiple first training samples from the multiple first data samples, each of the multiple first training samples preserving the privacy of the confidential data in the multiple first data samples by:
• (a) selecting a first subset of the multiple first data samples such that the first subset comprises data samples with identical first feature values, and
• (b) combining the first feature values of the first data samples of the first subset based on the first label of each of the first data samples of the first subset, by determining a weighted sum of the first feature values of the first data samples of the first subset wherein the feature value of a feature of the training sample is the weighted sum of the first feature values of that feature of the first data samples of the first subset;
• wherein the weighted sum comprises a sum of first feature values of that feature multiplied by the respective first labels of each of the data samples of the first subset,
• a second data collection device configured to:
• store the confidential data in the second data set comprising multiple second data samples, each of the multiple second data samples comprising multiple second feature values and a second label that classifies that second data sample;
• create second non-confidential training data by determining each of multiple second training samples from the multiple second data samples, each of the multiple second training samples preserving the privacy of the confidential data in the multiple second data samples by:
• (a) selecting a second subset of the multiple second data samples such that the second subset comprises data samples with identical second feature values, and
• (b) combining the second feature values of the second data samples of the second subset based on the second label of each of the second data samples of the second subset, by determining a weighted sum of the second feature values of the second data samples of the second subset wherein the feature value of a feature of the training sample is the weighted sum of the second feature values of that feature of the second data samples of the second subset, wherein the weighted sum comprises a sum of second feature values of that feature multiplied by the respective second labels of each of the data samples of the second subset;
• a data processing device configured to:
• merge each of the multiple first training samples with one of the multiple second training samples that is based on the same identical feature values as that first training sample to create merged non-confidential training data;
• train a classifier on the confidential data from the first data set and the second data set without access to the confidential data and with access to the merged non-confidential training data by calculating a classifier weight associated with a feature index from the merged non-confidential training data;
• and
• classify, using the trained classifier, a test value by determining a classification of the test value based on the classifier weight.