System and method for heterogeneous transferred learning for enhanced cybersecurity threat detection
A method includes training a first machine learning model with a first dataset, to produce a first trained machine learning model to infer cybersecurity-oriented file properties and/or detect cybersecurity threats within a first domain. The first dataset includes labeled files associated with the first domain. The first trained machine learning model includes multiple layers, some of which are trainable. A second trained machine learning model is generated, via a transfer learning process, using (1) at least one trainable layer from the multiple trainable layers of the first trained machine learning model, and (2) a second dataset different from the first dataset. The second dataset includes labeled files associated with a second domain. The first domain has a different syntax, different semantics, and/or a different structure than that of the second domain. The second trained machine learning model (e.g., a deep neural network model) is then available for use in inferring cybersecurity-oriented properties of the file in the second domain and/or detecting cybersecurity threats in the second domain.
1. A method for training machine learning models for cybersecurity threat detection, the method comprising:
training a first machine learning model with a first dataset including labeled files associated with a first domain, to produce a first trained machine learning model to infer cybersecurity-oriented file properties within the first domain, the first trained machine learning model including a plurality of trainable layers; and
generating a second trained machine learning model via a transfer learning process using (1) at least one trainable layer from the plurality of trainable layers of the first trained machine learning model, and (2) a second dataset different from the first dataset, the second dataset including labeled files associated with a second domain,
the first domain having at least one of a different syntax, different semantics, or different structure than that of the second domain.
2. The method of claim 1 , wherein each trainable layer from the plurality of trainable layers includes a matrix of weights.
3. The method of claim 1 , wherein at least one of the first dataset or the second dataset does not include data packets.
4. The method of claim 1 , further comprising selecting the first machine learning model for use in the transfer learning process, based on a detected overlap between the first dataset and the second dataset, prior to the generation of the second trained machine learning model.
5. The method of claim 1 , wherein the second dataset is smaller than the first dataset.
6. The method of claim 1 , wherein the second dataset includes an unrepresentative sample of subpopulations of the second domain.
7. The method of claim 1 , further comprising analyzing a labeled file associated with the second domain, using the second trained machine learning model, to determine a cybersecurity risk associated with the labeled file.
8. The method of claim 1 , further comprising selecting the at least one trainable layer from the plurality of trainable layers of the first trained machine learning model based on a resource constraint of a compute device.
9. A method for training machine learning models for cybersecurity threat detection, the method comprising:
receiving, at a machine learning system associated with a second domain, a first trained machine learning model to infer cybersecurity-oriented file properties within a first domain different from the second domain;
training the first trained machine learning model via a transfer learning process using a dataset including labeled files associated with the second domain to produce a second trained machine learning model; and
analyzing a labeled file associated with the second domain, at the machine learning system and using the second trained machine learning model, to determine a cybersecurity risk associated with the labeled file.
10. The method of claim 9 , wherein the first trained machine learning model includes a plurality of trainable layers, each trainable layer from the plurality of trainable layers including a matrix of weights.
11. The method of claim 9 , wherein the dataset does not include data packets.
12. The method of claim 9 , wherein the dataset is a second dataset, the method further comprising selecting the first machine learning model for use in the transfer learning process, based on a detected overlap between (1) a first dataset associated with the first trained machine learning model and (2) the second dataset, prior to the generation of the second trained machine learning model.
13. The method of claim 9 , wherein the dataset is a second dataset that is smaller than a first dataset associated with the first trained machine learning model.
14. The method of claim 9 , wherein the dataset includes an unrepresentative sample of subpopulations of the second domain.
15. The method of claim 9 , further comprising:
detecting that the cybersecurity risk has a value above a specified threshold; and
transmitting a signal representing an alert in response to detecting that the cybersecurity risk has the value above the specified threshold.
16. A method for training machine learning models for cybersecurity threat detection, the method comprising:
training a first machine learning model with a first dataset including labeled files associated with a first domain, to produce a first trained machine learning model to infer cybersecurity-oriented file properties within the first domain; and
generating a second trained machine learning model via a transfer learning process by:
modifying the first trained machine learning model to produce a modified first trained machine learning model, and
training the modified first trained machine learning model with a second dataset different from the first dataset, the second dataset including labeled files associated with a second domain, to produce the second trained machine learning model,
the first domain having at least one of a different syntax, different semantics, or different structure than that of the second domain.
17. The method of claim 16 , wherein the first trained machine learning model includes a plurality of trainable layers, each trainable layer from the plurality of trainable layers including a matrix of weights.
18. The method of claim 16 , wherein at least one of the first dataset or the second dataset does not include data packets.
19. The method of claim 16 , wherein the second dataset is smaller than the first dataset.
20. The method of claim 16 , wherein the second dataset includes an unrepresentative sample of subpopulations of the second domain.
21. The method of claim 16 , further comprising analyzing a labeled file associated with the second domain, using the second trained machine learning model, to determine a cybersecurity risk associated with the labeled file.
22. The method of claim 16 , wherein the modifying the first trained machine learning model is based on a resource constraint of a compute device.
23. The method of claim 16 , further comprising selecting the first machine learning model for use in the transfer learning process, based on a detected overlap between the first dataset and the second dataset, prior to the generation of the second trained machine learning model.