Granted Patent B2

US 12,235,990 · App. 17/751,397 · Granted Feb 25, 2025

Inventors: Martin Haerterich (Wiesloch, DE); Benjamin Weggenmann (Karlsruhe, DE)

Assignee: SAP SE

G06F21/6245

Abstract

Claims (61)

1. A system for obscuring private information in input data, the system comprising:

a computing device comprising at least one processor and an associated storage device, the at least one processor programmed to perform operations comprising:

accessing an input data unit;

applying an encoder model to the input data unit to generate a latent space representation of the input data unit in a latent space, the latent space having a first latent space dimension and a second latent space dimension;

applying multi-dimensional noise to the latent space representation of the input data unit, the multi-dimensional noise having a first value in the first latent space dimension and a second value different than the first value in the second latent space dimension; and

applying a decoder model to the latent space representation of the input data unit to generate an obscured data unit.

2. The system of claim 1 , the operations further comprising:

accessing a plurality of labeled input data units, a first labeled input data unit of the plurality of labeled input data units comprising a first input data unit feature, a first utility parameter describing the first input data unit feature, and a first privacy parameter describing the first input data unit feature;

generating, using the plurality of labeled input data units, a utility classifier, the utility classifier to map from the latent space to a corresponding utility; and

generating, using the plurality of labeled input data units, a privacy classifier, the privacy classifier to map from the latent space to a corresponding privacy.

3. The system of claim 1 , the operations further comprising:

determining, using a utility classifier, a first latent space dimension utility attribution label;

determining, using a privacy classifier, a first latent space dimension privacy attribution label; and

determining the first value of the multi-dimensional noise using the first latent space dimension utility attribution label and the first latent space dimension privacy attribution label.

4. The system of claim 3 , the operations further comprising:

generating a privacy explainability map using the input data unit and the privacy classifier, the privacy explainability map comprising a plurality of latent space dimension privacy attribution labels including the first latent space dimension utility attribution label; and

generating a utility explainability map using the input data unit and the utility classifier, the utility explainability map comprising a plurality of latent space dimension utility attribution labels including the first latent space dimension utility attribution label.

5. The system of claim 3 , the operations further comprising:

determining a ratio using the first latent space dimension utility attribution label and the first latent space dimension privacy attribution label; and

applying the ratio to a noise distribution, the first value of the multi-dimensional noise being based at least in part on the applying of the ratio to the noise distribution.

6. The system of claim 5 , the applying of the ratio to the noise distribution comprising applying the ratio to a variance of a Gaussian distribution.

7. The system of claim 5 , further comprising cropping a result of applying the ratio to the noise distribution to determine the first value of the multi-dimensional noise.

8. The system of claim 1 , the operations further comprising determining a plurality of vectors, the plurality of vectors comprising a first vector corresponding to the first latent space dimension and a second vector corresponding to the second latent space dimension.

9. A method for obscuring private information in input data, the method comprising:

accessing, by an obscuring system, an input data unit, the obscuring system comprising at least one processor and an associated storage device;

applying, by the obscuring system, an encoder model to the input data unit to generate a latent space representation of the input data unit in a latent space, the latent space having a first latent space dimension and a second latent space dimension;

applying, by the obscuring system, multi-dimensional noise to the latent space representation of the input data unit, the multi-dimensional noise having a first value in the first latent space dimension and a second value different than the first value in the second latent space dimension; and

applying, by the obscuring system, a decoder model to the latent space representation of the input data unit to generate an obscured data unit.

10. The method of claim 9 , further comprising:

accessing, by the obscuring system, a plurality of labeled input data units, a first labeled input data unit of the plurality of labeled input data units comprising a first input data unit feature, a first utility parameter describing the first input data unit feature, and a first privacy parameter describing the first input data unit feature;

generating, by the obscuring system and using the plurality of labeled input data units, a utility classifier, the utility classifier to map from the latent space to a corresponding utility; and

generating, by the obscuring system and using the plurality of labeled input data units, a privacy classifier, the privacy classifier to map from the latent space to a corresponding privacy.

11. The method of claim 9 , further comprising:

determining, using a utility classifier, a first latent space dimension utility attribution label;

determining, using a privacy classifier, a first latent space dimension privacy attribution label; and

determining the first value of the multi-dimensional noise using the first latent space dimension utility attribution label and the first latent space dimension privacy attribution label.

12. The method of claim 11 , further comprising:

generating, by the obscuring system, a privacy explainability map using the input data unit and the privacy classifier, the privacy explainability map comprising a plurality of latent space dimension privacy attribution labels including the first latent space dimension utility attribution label; and

generating, by the obscuring system, a utility explainability map using the input data unit and the utility classifier, the utility explainability map comprising a plurality of latent space dimension utility attribution labels including the first latent space dimension utility attribution label.

13. The method of claim 11 , further comprising:

determining, by the obscuring system, a ratio using the first latent space dimension utility attribution label and the first latent space dimension privacy attribution label; and

applying the ratio to a noise distribution by the obscuring system, the first value of the multi-dimensional noise being based at least in part on the applying of the ratio to the noise distribution.

14. The method of claim 13 , the applying of the ratio to the noise distribution comprising applying the ratio to a variance of a Gaussian distribution.

15. The method of claim 13 , further comprising cropping a result of applying the ratio to the noise distribution to determine the first value of the multi-dimensional noise.

16. The method of claim 9 , further comprising determining, by the obscuring system, a plurality of vectors, the plurality of vectors comprising a first vector corresponding to the first latent space dimension and a second vector corresponding to the second latent space dimension.

17. A machine-readable medium comprising instructions thereon that, when executed by at least one processor, cause the at least one processor to perform operations comprising:

accessing an input data unit;

applying an encoder model to the input data unit to generate a latent space representation of the input data unit in a latent space, the latent space having a first latent space dimension and a second latent space dimension;

applying multi-dimensional noise to the latent space representation of the input data unit, the multi-dimensional noise having a first value in the first latent space dimension and a second value different than the first value in the second latent space dimension; and

applying a decoder model to the latent space representation of the input data unit to generate an obscured data unit.

18. The machine-readable medium of claim 17 , the operations further comprising:

accessing a plurality of labeled input data units, a first labeled input data unit of the plurality of labeled input data units comprising a first input data unit feature, a first utility parameter describing the first input data unit feature, and a first privacy parameter describing the first input data unit feature;

generating, using the plurality of labeled input data units, a utility classifier, the utility classifier to map from the latent space to a corresponding utility; and

generating, using the plurality of labeled input data units, a privacy classifier, the privacy classifier to map from the latent space to a corresponding privacy.

19. The machine-readable medium of claim 17 , the operations further comprising:

determining, using a utility classifier, a first latent space dimension utility attribution label;

determining, using a privacy classifier, a first latent space dimension privacy attribution label; and

determining the first value of the multi-dimensional noise using the first latent space dimension utility attribution label and the first latent space dimension privacy attribution label.

20. The machine-readable medium of claim 19 , the operations further comprising:

generating a privacy explainability map using the input data unit and the privacy classifier, the privacy explainability map comprising a plurality of latent space dimension privacy attribution labels including the first latent space dimension utility attribution label; and

generating a utility explainability map using the input data unit and the utility classifier, the utility explainability map comprising a plurality of latent space dimension utility attribution labels including the first latent space dimension utility attribution label.

Assignments (1)

Continuity (1)

References Cited (46)

Cited By (1)

US 12,437,105