Neural network verification

1. A non-transitory computer-readable medium storing computer-executable instructions that when executed by a computer cause the computer to:

train a neural network with a training data set to perform a predefined task;

train the neural network with a sentinel data set, where the sentinel data set causes the neural network to provide an identification signal in response to a predefined query set;

provide the predefined query set, the identification signal, and access to the neural network to a trusted third party, where the trusted third party confirms whether the predefined query set, the identification signal, and the neural network conflict with an archived query set, an archived identification signal, and an archived neural network, and where the trusted third party confirms whether the neural network provides the identification signal in response to the predefined query set;

verify whether a suspicious service operates an unauthorized copy of the neural network by extracting the identification signal from responses the suspicious service provides upon receiving the predefined query set; and

receive a confirmation from the trusted third party when the suspicious service operates the unauthorized copy of the neural network.

2. The non-transitory computer-readable medium of claim 1 , where the instructions further cause the computer to:

enhance the identification signal with an error correcting code, where the error correcting code inhibits modifications to the unauthorized copy of the neural network by causing the suspicious service to provide a non-matching identification signal in response to the predefined query set.

3. The non-transitory computer-readable medium of claim 1 , where extracting the identification signal comprises hiding the predefined query set within a masking query set to prevent an operator of the suspicious service from identifying the predefined query set.

4. The non-transitory computer-readable medium of claim 1 , where the sentinel data set is uncorrelated with the training data set.

5. The non-transitory computer-readable medium of claim 4 , where the sentinel data set comprises randomly generated inputs.

6. The non-transitory computer-readable medium of claim 1 , where the neural network topology is designed to accommodate the sentinel data set to prevent under-fitting the neural network.

7. The non-transitory computer-readable medium of claim 1 , where the trusted third party verifies whether the suspicious service operates the unauthorized copy of the neural network by extracting the identification signal from responses the suspicious service provides to the predefined query set.

8. A computing system, comprising:

a processor;

a training logic to train a neural network according to a training data set to perform a predefined task;

a neural network security logic to train the neural network according to a first sentinel data set, creating a first secured neural network, where the first sentinel data set causes the first secured neural network to provide a first identification signal in response to a first predefined query set;

logic to provide the first predefined query set, the first identification signal, and access to the neural network to a trusted third party, where the trusted third party confirms whether the predefined query set, the identification signal, and the neural network conflict with an archived query set, an archived identification signal, and an archived neural network, and where the trusted third party confirms whether the neural network provides the identification signal in response to the predefined query set;

an embedding logic to embed the first secured neural network into a first unsecured device;

a verification logic to identify whether a suspicious service operates an unauthorized copy of the neural network by extracting the first identification signal from responses the suspicious service provides to the first predefined query set; and

logic to receive a confirmation from the trusted third party when the suspicious service operates the unauthorized copy of the neural network.

9. The system of claim 8 ,

where the neural network security logic also trains the neural network according to a second sentinel data set, creating a second secured neural network, where the second sentinel data set causes the second secured neural network to provide a second identification signal in response to a second predefined query set;

where the embedding logic embeds the second secured neural network into a second unsecured device; and

where the verification logic identifies whether the suspicious service operates the unauthorized copy of the neural network by extracting the second identification signal from a response the suspicious service provides to the second predefined query set.

10. The system of claim 9 , where the suspicious service providing the first identification signal in response to the first predefined query set indicates the unauthorized copy of the neural network was copied from the first unsecured device, and where the suspicious service providing the second identification signal in response to the second predefined query set indicates the unauthorized copy of the neural network was copied from the second unsecured device.

11. The system of claim 8 , comprising an error correction logic to embed an error correcting code into the first secured neural network, where the error correcting code inhibits modifications to the unauthorized copy of the neural network that affect the provision of the first identification signal in response to the first predefined query set.

12. The system of claim 8 , where the first unsecured device is associated with one or more of a person, an instance of a service, and a company.