# US 12,769,845 B2 · Systems and methods for measuring data exfiltration vulnerability and dynamic differential privacy in a zero-trust computing environment

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,769,845 B2
- Document ID: US12769845B2
- Status: Granted
- Kind: B2
- Application number: 18/110,308
- Filed: 2023-02-15
- Published: 2024-08-15
- Granted: 2026-10-20
- Earliest claimed: 2023-02-15
- Assignees: BeeKeeperAI, Inc.
- Applicants: BeeKeeperAI, Inc.
- Inventors: Mary Elizabeth Chalk; Robert Derward Rogers; Alan Donald Czeszynski
- Web page: https://patentgrep.com/patent/US12769845B2

## CPC

- G06F 21/577: {Assessing vulnerabilities and evaluating computer system security}
- G06F 21/6245: {Protecting personal data, e.g. for financial or medical purposes}
- G06N 20/00: Machine learning
- G06F 2221/033: Test or assess software
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

An algorithm is trained on a known dataset to facilitate dynamic data exfiltration protection in a zero-trust environment. The classifications generated by the trained algorithm on a very large set of inputs may then be used to train an inversion threat model by a bad actor attempting to exfiltrate data from the data steward. Since our system is taking place within the enclave/secure computing node, the system is able to very accurately build an inversion threat model since the original training dataset is known (a ‘gold standard’ inversion model). This inversion model can be characterized to determine its performance/accuracy of properly identifying a given input as being within the original training dataset or not (a data exfiltration event). This very accurate inversion model will be superior at data exfiltration as compared to any inversion attack model generated by a bad actor using only the algorithm classification outputs. As such, the results of this inversion model provide a ceiling on the likelihood of data exfiltration. Very accurate inversion models indicate that data exfiltration is easier/more likely. A poor performing inversion model indicates that the training data is more secure and less able to be exfiltrated. Differential privacy may be fluctuated to manage the risk of data exfiltration.

## Claims

### Claim 1

A computerized method of data exfiltration analysis in a zero-trust computing environment, the method comprising:
receiving a data set within a secure computing node;
splitting the data set into a plurality of orthogonal subsets;
training a plurality of models on each of the plurality of orthogonal subsets;
combining each of the models using a weighted average into an algorithm;
iteratively training the algorithm on the data set to generate a set of weights, wherein each iteration of the training improves model performance, and wherein the iterations are concluded when the model performance improvement between a current model and a prior model is below an improvement threshold, and wherein a final model generated after the iterations is a trained algorithm;
training an inversion model on the trained algorithm;
determining a degree of sensitivity for the data set;
calculating an upper limit of practical data exfiltration based upon inversion model complexity and algorithm complexity and input availability
characterizing the performance of the inversion model;
setting a dynamic threshold that is proportional to the degree of sensitivity of the data set and the upper limit of practical data exfiltration;
comparing the inversion model performance against the dynamic threshold; and
injecting noise into the data set and retraining the algorithm on the noise injected data set and repeating the noise injection until the inversion model performance is below the dynamic threshold.

### Claim 2 (depends on 1)

The method of claim 1, wherein the threshold is configurable.

### Claim 3 (depends on 2)

The method of claim 2, wherein the threshold is configured by at least one of a regulator and a data steward.

### Claim 4 (depends on 2)

The method of claim 2, wherein the threshold is configured based upon a sensitivity level of the data set.

### Claim 5 (depends on 1)

The method of claim 1, wherein the noise is Gaussian noise.

### Claim 6 (depends on 1)

The method of claim 1, wherein the noise is applied to only a portion of the data set.

### Claim 7 (depends on 6)

The method of claim 6, wherein the portion of the data set is data determined to be sensitive.

### Claim 8 (depends on 1)

The method of claim 1, wherein the noise is generated by the inversion model.

### Claim 9 (depends on 1)

The method of claim 1, further comprising when the inversion model performance is below the threshold outputting the algorithm weights.

### Claim 10

A zero-trust computer system for data exfiltration analysis comprising:
a database for receiving a data set within a secure computing node;
a runtime server, comprising a processor, for splitting the data set into a plurality of orthogonal subsets, training a plurality of models on each of the plurality of orthogonal subsets, combining each of the models using a weighted average into an algorithm, iteratively training the algorithm on the data set to generate a set of weights, wherein each iteration of the training improves model performance, and wherein the iterations are concluded when the model performance improvement between a current model and a prior model is below an improvement threshold, and wherein a final model generated after the iterations is a trained algorithm;
the runtime server further configured for determining a degree of sensitivity for the data set, and calculating an upper limit of practical data exfiltration based upon inversion model complexity and algorithm complexity and input availability; and
a threat vulnerability server, comprising a processor, for training an inversion model on the trained algorithm, setting a dynamic threshold that is proportional to the degree of sensitivity of the data set and the upper limit of practical data exfiltration characterizing the performance of the inversion model, comparing the inversion model performance against the dynamic threshold, and inject noise into the data set and retrain the algorithm on the noise injected data set and repeating the noise injection until the inversion model performance is below the dynamic threshold.

### Claim 11 (depends on 10)

The zero-trust computer system of claim 10, wherein the threshold is configurable.

### Claim 12 (depends on 11)

The zero-trust computer system of claim 11, wherein the threshold is configured by at least one of a regulator and a data steward.

### Claim 13 (depends on 11)

The zero-trust computer system of claim 11, wherein the threshold is configured based upon a sensitivity level of the data set.

### Claim 14 (depends on 10)

The zero-trust computer system of claim 10, wherein the noise is Gaussian noise.

### Claim 15 (depends on 10)

The zero-trust computer system of claim 10, wherein the noise is applied to only a portion of the data set.

### Claim 16 (depends on 15)

The zero-trust computer system of claim 15, wherein the portion of the data set is data determined to be sensitive.

### Claim 17 (depends on 10)

The zero-trust computer system of claim 10, wherein the noise is generated by the inversion model.

### Claim 18 (depends on 10)

The zero-trust computer system of claim 10, wherein the threat vulnerability server is further configured to output the algorithm weights when the inversion model performance is below the threshold.