Methods and apparatus for extracting data in deep neural networks

1. A method of extracting data in a deep learning model, the method comprising:

receiving an input query;

determining a first decision boundary set, wherein the first decision boundary is a subset of a decision boundary set corresponding to a target layer of the deep learning model, wherein the first decision boundary set surrounds a decision region, wherein the determining of the first decision boundary set comprises:

determining a distribution probability of decision boundaries included in the decision boundary set, wherein determining the distribution probability of the decision boundaries comprises optimizing the distribution probability of the decision boundaries by minimizing a loss function corresponding to the target layer; and

determining the first decision boundary set based on the distribution probability of the decision boundaries;

extracting the decision region, wherein the decision region comprises the input query, the extracting based on the first decision boundary set; and

extracting data included in the decision region to determine if the decision region is properly trained, the determining comprising:

determining that a data sample in the extracted data is unexpected; and

based on the determining, retraining the decision region, wherein the retraining narrows the first decision boundary set and the decision area, wherein the narrowed decision region comprises more robust data samples than the decision regions.

2. The method of claim 1 , wherein retraining the decision region comprises re-training the deep learning model based on the extracted data.

3. The method of claim 1 , wherein the determining of the first decision boundary set based on the distribution probability of the decision boundaries comprises determining decision boundaries of which a distribution probability is greater than or equal to a threshold, among the decision boundaries, to be the first decision boundary set.

4. The method of claim 1 , wherein determining the distribution probability of the decision boundaries included in the decision boundary set comprises determining Bernoulli parameters of the decision boundaries.

5. The method of claim 4 , wherein minimizing the loss function corresponding to the target layer comprises optimizing the Bernoulli parameters.

6. The method of claim 1 , wherein the extracting of the data comprises extracting the data based on a rapidly-exploring random tree (RRT) algorithm having the decision region as a constraint.

7. The method of claim 1 , wherein a shape of the region is a cube shape.

8. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, causes the processor to perform the method, the method comprising:

receiving an input query;

determining a first decision boundary set, wherein the first decision boundary is a subset of a decision boundary set corresponding to a target layer of the deep learning model, wherein the first decision boundary set surrounds a decision region, wherein the determining of the first decision boundary set comprises:

determining a distribution probability of decision boundaries included in the decision boundary set, wherein determining the distribution probability of the decision boundaries comprises optimizing the distribution probability of the decision boundaries by minimizing a loss function corresponding to the target layer; and

determining the first decision boundary set based on the distribution probability of the decision boundaries;

extracting the decision region, wherein the decision region comprises the input query, the extracting based on the first decision boundary set; and

extracting data included in the decision region to determine if the decision region is properly trained, the determining comprising:

determining that a data sample in the extracted data is unexpected; and

based on the determining, retraining the decision region, wherein the retraining narrows the first decision boundary set and the decision area, wherein the narrowed decision region comprises more robust data samples than the decision regions.

9. An apparatus for extracting data in a deep learning model, the apparatus comprising:

a processor configured to receive an input query, determine a first decision boundary set being a subset of a decision boundary set corresponding to a target layer of the deep learning model, extract a decision region, wherein the decision region comprises the input query, the extracting based on the first decision boundary set, and extract data included in the decision region,

wherein the processor is configured to determine a distribution probability of decision boundaries included in the decision boundary set, and determine the first decision boundary set based on the distribution probability of the decision boundaries, wherein determining the distribution probability of the decision boundaries comprises optimizing the distribution probability of the decision boundaries by minimizing a loss function corresponding to the target layer,

wherein extracting data included in the decision region comprises determining if the decision region is properly trained, the determining comprising:

determining that a data sample in the extracted data is unexpected; and

based on the determining, retraining the decision region, wherein the retraining narrows the first decision boundary set and the decision area, wherein the narrowed decision region comprises more robust data samples than the decision regions.

10. The apparatus of claim 9 , wherein retraining the decision region comprises retraining the deep learning model based on the extracted data.

11. The apparatus of claim 9 , wherein the processor is configured to determine decision boundaries of which a distribution probability is greater than or equal to a threshold, among the decision boundaries, to be the first decision boundary set.

12. The apparatus of claim 9 , wherein the processor is configured to determine Bernoulli parameters of the decision boundaries.

13. The apparatus of claim 12 , wherein the processor is configured to minimize the loss function corresponding to the target layer by optimizing the Bernoulli parameters.

14. The apparatus of claim 12 , wherein the processor is configured to extract the data based on a rapidly-exploring random tree (RRT) algorithm having the decision region as a constraint.

15. The apparatus of claim 9 , wherein a shape of the region is a cube shape.