Privacy preserving ensemble learning as a service
Techniques described herein relate to a method for predicting results using ensemble models. The method may include receiving trained model data sets from a model source nodes, each trained model data set comprising a trained model, an important feature list, and a missing feature generator; receiving a prediction request data set; making a determination that the prediction request data set does not include an input feature for a trained model; generating, based on the determination and using a missing feature generator, a substitute feature to replace the input feature; executing the trained model using the prediction request data set and the substitute feature to obtain a first prediction; executing a second trained model using the prediction request data set to obtain a second prediction; and obtaining a final prediction using the first prediction, the second prediction, and an ensemble model.
1. A method for predicting results using ensemble models, the method comprising:
receiving a first trained model data set from a first model source node, the first trained model data set comprising:
a first trained model,
a first critical feature list comprising critical properties of the first trained model data set, wherein the critical properties of the first training model data set are inputs for the first trained model relevant to a result produced by the first trained model, and
a first missing feature generator configured to generate missing critical properties of the first trained model data set based on the first critical feature list;
receiving a second trained model data set from a second model source node, the second trained model data set comprising:
a second trained model,
a second critical feature list comprising critical properties of the second trained model data set, wherein the critical properties of the second trained model data set are inputs for the second trained model relevant to a result produced by the second trained model, and
a second missing feature generator, configured to generate missing critical properties of the second trained model data set based on the second critical feature list;
receiving a prediction request data set comprising a third critical feature list, the third critical feature list comprising critical properties of the prediction request data set, wherein the critical properties of the prediction request data set are properties to be compared to the first critical feature list and the second critical feature list;
making a first determination that the third critical feature list is missing at least one critical feature from the first critical feature list;
generating, based on the first determination and using the first missing feature generator and using a General Adversarial Network (GAN), a first substitute feature to replace the at least one missing critical feature, to obtain a modified prediction request data set;
executing the first trained model using the modified prediction request data set to obtain a first prediction;
executing the second trained model using the prediction request data set to obtain a second prediction; and
obtaining a final prediction using the first prediction and the second prediction as inputs to an ensemble model.
2. The method of claim 1 , further comprising, before receiving the prediction request data set, training the ensemble model using an ensemble model training set and a plurality of missing feature values generated using a plurality of missing feature generators.
3. The method of claim 1 , wherein the first missing feature generator is a GAN.
4. The method of claim 1 , further comprising, before generating the first substitute feature, making a second determination that the at least one missing critical feature is not included in the first critical feature list, wherein generating the first substitute feature is further based on the second determination.
5. The method of claim 1 , wherein the prediction request data set is received from an entity seeking a prediction related to a potential cyber-attack.
6. The method of claim 1 , wherein:
the first trained model data set is received from a first participant entity in a collaborative ensemble modeling service,
the second trained model data set is received from a second participant entity in a collaborative ensemble modeling service, and
no data used to train the first trained model and the second trained model is shared between the first participant entity and the second participant entity.
7. The method of claim 1 , wherein the final prediction is obtained, at least in part, using a weighted average of the first prediction and the second prediction.
8. A non-transitory computer readable medium comprising computer readable program code, which when executed by a computer processor enables the computer processor to perform a method for predicting results using ensemble models, the method comprising:
receiving a first trained model data set from a first model source node, the first trained model data set comprising:
a first trained model,
a first critical feature list comprising critical properties of the first trained model data set, wherein the critical properties of the first training model data set are inputs for the first trained model relevant to a result produced by the first trained model, and
a first missing feature generator configured to generate missing critical properties of the first trained model data set based on the first critical feature list;
receiving a second trained model data set from a second model source node, the second trained model data set comprising:
a second trained model,
a second critical feature list comprising critical properties of the second trained model data set, wherein the critical properties of the second trained model data set are inputs for the second trained model relevant to a result produced by the second trained model, and
a second missing feature generator, configured to generate missing critical properties of the second trained model data set based on the second critical feature list;
receiving a prediction request data set comprising a third critical feature list, the third critical feature list, comprising critical properties of the prediction request data set, wherein the critical properties of the prediction request data set are properties to be compared to the first critical feature list and the second critical feature list;
making a first determination that the third critical feature list is missing at least one critical feature from the first critical feature list;
generating, based on the first determination and using the first missing feature generator and using a General Adversarial Network (GAN), a first substitute feature to replace the at least one missing critical feature, to obtain a modified prediction request data set;
executing the first trained model using the modified prediction request data set to obtain a first prediction;
executing the second trained model using the prediction request data set to obtain a second prediction; and
obtaining a final prediction using the first prediction and the second prediction as inputs to an ensemble model.
9. The non-transitory computer readable medium of claim 8 , wherein the method performed by executing the computer readable program code further comprises, before receiving the prediction request data set, training the ensemble model using an ensemble model training set and a plurality of missing feature values generated using a plurality of missing feature generators.
10. The non-transitory computer readable medium of claim 8 , wherein the first missing feature generator is a GAN.
11. The non-transitory computer readable medium of claim 8 , wherein the method performed by executing the computer readable program code further comprises, before generating the first substitute feature, making a second determination that the at least one missing critical feature is not included in the first critical feature list, wherein generating the first substitute feature is further based on the second determination.
12. The non-transitory computer readable medium of claim 8 , wherein the prediction request data set is received from an entity seeking a prediction related to a potential cyber-attack.
13. The non-transitory computer readable medium of claim 8 , wherein:
the first trained model data set is received from a first participant entity in a collaborative ensemble modeling service,
the second trained model data set is received from a second participant entity in a collaborative ensemble modeling service, and
no data used to train the first trained model and the second trained model is shared between the first participant entity and the second participant entity.
14. The non-transitory computer readable medium of claim 8 , wherein the final prediction is obtained, at least in part, using a weighted average of the first prediction and the second prediction.
15. A system for predicting results using ensemble models, the system comprising:
a model aggregator comprising memory and a processor comprising circuitry, operatively connected to a plurality of model source nodes, and configured to:
receiving a first trained model data set from a first model source node, the first trained model data set comprising:
a first trained model,
a first critical feature list comprising critical properties of the first trained model data set, wherein the critical properties of the first trained model data set are inputs for the first trained model relevant to a result produced by the first trained model, and
a first missing feature generator configured to generate missing critical properties of the first trained model data set based on the first critical feature list;
receiving a second trained model data set from a second model source node, the second trained model data set comprising:
a second trained model,
a second critical feature list comprising critical properties of the second trained model data set, wherein the critical properties of the second trained model data set are inputs for the second trained model relevant to a result produced by the second trained model, and
a second missing feature generator, configured to generate missing critical properties of the second trained model data set based on the second critical feature list;
receiving a prediction request data set comprising a third critical feature list, the third critical feature list, comprising critical properties of the prediction request data set, wherein the critical properties of the prediction request data set are properties to be compared to the first critical feature list and the second critical feature list;
making a first determination that the third critical feature list is missing at least one critical feature from the first critical feature list;
generating, based on the first determination and using the first missing feature generator and using a General Adversarial Network (GAN), a first substitute feature to replace the at least one missing critical feature, to obtain a modified prediction request data set;
executing the first trained model using the modified prediction request data set to obtain a first prediction;
executing the second trained model using the prediction request data set to obtain a second prediction; and
obtaining a final prediction using the first prediction and the second prediction as inputs to an ensemble model.
16. The system of claim 15 , wherein the model aggregator is further configured to, before receiving the prediction request data set, train the ensemble model using an ensemble model training set and a plurality of missing feature values generated using a plurality of missing feature generators.
17. The system of claim 15 , wherein the first missing feature generator is a GAN.
18. The system of claim 15 , wherein the model aggregator is further configured to, before generating the first substitute feature, make a second determination that the at least one missing critical feature is not included in the first critical feature list, wherein generating the first substitute feature is further based on the second determination.
19. The system of claim 15 , wherein the prediction request data set is received from an entity seeking a prediction related to a potential cyber-attack.
20. The system of claim 15 , wherein:
the first trained model data set is received from a first participant entity in a collaborative ensemble modeling service,
the second trained model data set is received from a second participant entity in a collaborative ensemble modeling service, and
no data used to train the first trained model and the second trained model is shared between the first participant entity and the second participant entity.