Scanning and detecting threats in machine learning models
A machine learning model is scanned to detect actual or potential threats. The threats can be detected before execution of the machine learning model or during an isolated execution environment. The threat detection may include performing a machine learning file format check, vulnerability check, tamper check, and stenography check. The machine learning model may also be monitored in an isolated environment during an execution or runtime session. After performing a scan, the system can generate a signature based on actual, potential, or absence of detected threats.
1. A method for scanning a machine learning model for threats, comprising:
receiving, by a scanning module, data for a machine learning model, the scanning module stored on a first server, the data associated with model parameters and received before execution of the machine learning model;
performing, by the scanning module, a plurality of checks based on the received machine learning model data, the checks performed while the machine learning model is not executing,
the performing comprising:
determining weights and biases of the machine learning model;
determining an expected entropy for the machine learning model based on clustering of known machine learning models guided by calculated entropy;
determining, based on the weights and biases, an actual entropy for the machine learning model; and
calculating the difference between the expected entropy and the actual entropy;
identifying, by the scanning module and based on the calculated difference, whether the machine learning model includes a threat within the machine learning model based on results of the plurality of checks; and
adding, by the scanning module and based on the identifying, an indicator to the machine learning model characterizing actual, potential or detected threats within the machine learning model.
2. The method of claim 1 , further including:
generating a signature based on determining that the machine learning model does not include a threat; and
embedding the generated signature within the data for the machine learning model.
3. The method of claim 1 , wherein the plurality of checks includes two or more of a file format check, a vulnerability check, a tampering check, and a stenography check.
4. The method of claim 1 , further including:
executing the machine learning model in an isolated environment;
monitoring the execution of the machine learning in the isolated environment to detect suspicious activity.
5. The method of claim 1 , wherein an identified threat may be a potential threat or an actual threat.
6. The method of claim 1 , wherein the first server which stores the scanning module is within a computing environment that includes the machine learning model.
7. A non-transitory computer readable storage medium having embodied thereon a program, the program being executable by a processor to perform a method for scanning a machine learning model for threats, the method comprising:
receiving, by a scanning module, data for a machine learning model, the scanning module stored on a first server, the data associated with model parameters and received before execution of the machine learning model;
performing, by the scanning module, a plurality of checks based on the received machine learning model data, the checks performed while the machine learning model is not executing,
the performing comprising:
determining weights and biases of the machine learning model;
determining an expected entropy for the machine learning model based on clustering of known machine learning models guided by calculated entropy;
determining, based on the weights and biases, an actual entropy for the machine learning model; and
calculating the difference between the expected entropy and the actual entropy;
identifying, by the scanning module and based on the calculated difference, whether the machine learning model includes a threat within the machine learning model based on results of the plurality of checks;
adding, by the scanning module and based on the identifying, an indicator to the machine learning model characterizing actual, potential or detected threats within the machine learning model; and
reporting the results to a remote computing device by the scanning module.
8. The non-transitory computer readable storage medium of claim 7 , the method further including generating a signature based on determining that the machine learning model does not include a threat, and embedding the generated signature within the data for the machine learning model.
9. The non-transitory computer readable storage medium of claim 7 , wherein the plurality of checks includes two or more of a file format check, a vulnerability check, a tampering check, and a stenography check.
10. The non-transitory computer readable storage medium of claim 7 , the method further including executing the machine learning model in an isolated environment, and monitoring the execution of the machine learning in the isolated environment to detect suspicious activity.
11. The non-transitory computer readable storage medium of claim 7 , wherein the first server which stores the scanning module is within a computing environment that includes the machine learning model.
12. A system for scanning a machine learning model for threats, comprising:
one or more servers including a memory and a processor; and
wherein the memory stores one or more modules which, when executed by the processor, perform operations comprising:
receiving, by a scanning module, data for a machine learning model, the data associated with model parameters and received before execution of the machine learning model,
performing, by the scanning module, a plurality of checks based on the received machine learning model data, the checks performed while the machine learning model is not executing,
the performing comprising:
determining weights and biases of the machine learning model;
determining an expected entropy for the machine learning model based on clustering of known machine learning models guided by calculated entropy;
determining, based on the weights and biases, an actual entropy for the machine learning model; and
calculating the difference between the expected entropy and the actual entropy;
identifying, by the scanning module and based on the calculated difference, whether the machine learning model includes a threat within the machine learning model based on results of the plurality of checks, and
adding, by the scanning module and based on the identifying, an indicator to the machine learning model characterizing actual, potential or detected threats within the machine learning model; and
reporting the results to a remote computing device by the scanning module.
13. The system of claim 12 , the operations further comprise:
generating a signature based on determining that the machine learning model does not include a threat, and embedding the generated signature within the data for the machine learning model.
14. The system of claim 12 , wherein the plurality of checks includes two or more of a file format check, a vulnerability check, a tampering check, and a stenography check.
15. The system of claim 12 , the operations further comprise:
executing the machine learning model in an isolated environment, and monitoring the execution of the machine learning in the isolated environment to detect suspicious activity.
16. The method of claim 1 , wherein the identifying comprises:
determining, when the machine learning model is a fine-tuned version of a base model, whether the machine learning model has a same topology architecture as the base model;
wherein:
a determination that the machine learning model has the same topology architecture as the base model indicates that there are no threats within the machine learning model; and
a determination that the machine learning model has a different topology architecture as the base model indicates that there is a threat within the machine learning model.
17. The method of claim 1 , wherein the identifying comprises:
detecting whether there are dormant portions within the machine learning model which are activated upon a certain input;
wherein:
a determination that the machine learning model does not have dormant portions which are activated upon a certain input indicates that there are no threats within the machine learning model; and
a determination that the machine learning model has dormant portions which are activated upon a certain input indicates that there is a threat within the machine learning model.
18. The method of claim 1 , wherein the identifying comprises:
detecting whether a distribution of lower order bits within the model parameters of the machine learning model is random;
wherein:
a determination that the distribution of the lower order bits within the model parameters of the machine learning model is random indicates that there are no threats within the machine learning model; and
a determination that the distribution of the lower order bits within the model parameters of the machine learning model is not random indicates that there is a threat within the machine learning model.
19. The system of claim 12 , wherein the identifying comprises:
determining, when the machine learning model is a fine-tuned version of a base model, whether the machine learning model has a same topology architecture as the base model;
wherein:
a determination that the machine learning model has the same topology architecture as the base model indicates that there are no threats within the machine learning model; and
a determination that the machine learning model has a different topology architecture as the base model indicates that there is a threat within the machine learning model.
20. The system of claim 12 , wherein the identifying comprises:
detecting whether there are dormant portions within the machine learning model which are activated upon a certain input;
wherein:
a determination that the machine learning model does not have dormant portions which are activated upon a certain input indicates that there are no threats within the machine learning model; and
a determination that the machine learning model has dormant portions which are activated upon a certain input indicates that there is a threat within the machine learning model.
21. The system of claim 12 , wherein the identifying comprises:
detecting whether a distribution of lower order bits within the model parameters of the machine learning model is random;
wherein:
a determination that the distribution of the lower order bits within the model parameters of the machine learning model is random indicates that there are no threats within the machine learning model; and
a determination that the distribution of the lower order bits within the model parameters of the machine learning model is not random indicates that there is a threat within the machine learning model.