Deep learning-based detection of phishing links

1. A method comprising:

extracting, from a link contained in an electronic message received from an upstream device on a network, first unit-level input data of a first semantic type and second unit-level input data of a second semantic type;

in response to inputting the first and second unit-level input data into first and second deep learning models, respectively, outputting, by the first and second deep learning models, first and second unit-level classification data that corresponds to the first and second unit-level input data, respectively, the first deep learning model having been trained to recognize, in unit-level data of the first semantic type, first patterns of syntactic features and semantic features that are predictive of phishing and the second deep learning model having been trained to recognize, in unit-level data of the second semantic type, second patterns of syntactic features and semantic features that are predictive of phishing;

combining the first and second unit-level classification data with at least one numeric feature that has been extracted from context data to produce link-level input data;

in response to inputting the link-level input data into a third deep learning model, outputting, by the third deep learning model, link-level classification data, the third deep learning model having been trained to recognize patterns of syntactic features and semantic features and numeric features that are predictive of phishing;

in response to the link-level classification data matching a criterion, causing the network to modify, delay, or block transmission of the electronic message to a downstream device.

2. The method of claim 1 , further comprising performing the extracting of the first unit-level input data using a grammar to identify a sequence of characters in the link that corresponds to at least one domain name.

3. The method of claim 1 , further comprising performing the extracting of the second unit-level input data using a grammar to identify a sequence of characters in the link that corresponds to at least one path name.

4. The method of claim 2 , further comprising using the at least one domain name to determine Autonomous System Number (ASN) data associated with the domain name and using at least some of the ASN data as the at least one numeric feature.

5. The method of claim 4 , further comprising using the at least one domain name to determine geo-location data associated with the at least one domain name and using the geo-location data as the at least one numeric feature.

6. The method of claim 5 , further comprising creating the link-level input data by concatenating the first unit-level classification data and the second unit-level classification data and the ASN data and the geo-location data.

7. The method of claim 1 , the first and second deep learning models having been trained to classify sequences of characters that include bait text that is predictive of phishing differently than sequences of characters that do not include the bait text.

8. The method of claim 1 , the first and second deep learning models having been trained to classify different sequences of bait text and non-bait text and separators differently.

9. The method of claim 1 , the first and second and third deep learning models comprising artificial convolutional neural networks or layers of an artificial convolutional neural network.

10. The method of claim 1 , further comprising receiving the electronic message from a message transfer application on the network.

11. An apparatus comprising:

one or more non-transitory computer-readable storage media comprising instructions which, when executed by one or more processors, cause the one or more processors to perform operations comprising:

extracting, from a link contained in an electronic message received from an upstream device on a network, first unit-level input data of a first semantic type and second unit-level input data of a second semantic type;

in response to inputting the first and second unit-level input data into first and second deep learning models, respectively, outputting, by the first and second deep learning models, first and second unit-level classification data that corresponds to the first and second unit-level input data, respectively, the first deep learning model having been trained to recognize, in unit-level data of the first semantic type, first patterns of syntactic features and semantic features that are predictive of phishing and the second deep learning model having been trained to recognize, in unit-level data of the second semantic type, second patterns of syntactic features and semantic features that are predictive of phishing;

combining the first and second unit-level classification data with at least one numeric feature that has been extracted from context data to produce link-level input data;

in response to inputting the link-level input data into a third deep learning model, outputting, by the third deep learning model, link-level classification data, the third deep learning model having been trained to recognize patterns of syntactic features and semantic features and numeric features that are predictive of phishing;

in response to the link-level classification data matching a criterion, causing the network to modify, delay, or block transmission of the electronic message to a downstream device.

12. The apparatus of claim 11 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising performing the extracting of the first unit-level input data using a grammar to identify a sequence of characters in the link that corresponds to at least one domain name.

13. The apparatus of claim 11 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising performing the extracting of the second unit-level input data using a grammar to identify a sequence of characters in the link that corresponds to at least one path name.

14. The apparatus of claim 12 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising using the at least one domain name to determine Autonomous System Number (ASN) data associated with the domain name and using at least some of the ASN data as the at least one numeric feature.

15. The apparatus of claim 14 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising using the at least one domain name to determine geo-location data associated with the at least one domain name and using the geo-location data as the at least one numeric feature.

16. The apparatus of claim 15 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising creating the link-level input data by concatenating the first unit-level classification data and the second unit-level classification data and the ASN data and the geo-location data.

17. The apparatus of claim 11 , the first and second deep learning models having been trained to classify sequences of characters that include bait text that is predictive of phishing differently than sequences of characters that do not include the bait text.

18. The apparatus of claim 11 , the first and second deep learning models having been trained to classify different sequences of bait text and non-bait text and separators differently.

19. The apparatus of claim 11 , the first and second and third deep learning models comprising artificial convolutional neural networks or layers of an artificial convolutional neural network.

20. The apparatus of claim 11 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising receiving the electronic message from a message transfer application on the network.

21. A system comprising:

one or more processors;

one or more non-transitory storage media storing instructions which, when executed by the one or more processors, cause:

extracting, from a link contained in an electronic message received from an upstream device on a network, first unit-level input data of a first semantic type and second unit-level input data of a second semantic type;

in response to inputting the first and second unit-level input data into first and second deep learning models, respectively, outputting, by the first and second deep learning models, first and second unit-level classification data that corresponds to the first and second unit-level input data, respectively, the first deep learning model having been trained to recognize, in unit-level data of the first semantic type, first patterns of syntactic features and semantic features that are predictive of phishing and the second deep learning model having been trained to recognize, in unit-level data of the second semantic type, second patterns of syntactic features and semantic features that are predictive of phishing;

combining the first and second unit-level classification data with at least one numeric feature that has been extracted from context data to produce link-level input data;

in response to inputting the link-level input data into a third deep learning model, outputting, by the third deep learning model, link-level classification data, the third deep learning model having been trained to recognize patterns of syntactic features and semantic features and numeric features that are predictive of phishing;

in response to the link-level classification data matching a criterion, causing the network to modify, delay, or block transmission of the electronic message to a downstream device.

22. The system of claim 21 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising performing the extracting of the first unit-level input data using a grammar to identify a sequence of characters in the link that corresponds to at least one domain name.

23. The system of claim 21 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising performing the extracting of the second unit-level input data using a grammar to identify a sequence of characters in the link that corresponds to at least one path name.

24. The system of claim 22 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising using the at least one domain name to determine Autonomous System Number (ASN) data associated with the domain name and using at least some of the ASN data as the at least one numeric feature.

25. The system of claim 24 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising using the at least one domain name to determine geo-location data associated with the at least one domain name and using the geo-location data as the at least one numeric feature.

26. The system of claim 25 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising creating the link-level input data by concatenating the first unit-level classification data and the second unit-level classification data and the ASN data and the geo-location data.

27. The system of claim 21 , the first and second deep learning models having been trained to classify sequences of characters that include bait text that is predictive of phishing differently than sequences of characters that do not include the bait text.

28. The system of claim 21 , the first and second deep learning models having been trained to classify different sequences of bait text and non-bait text and separators differently.

29. The system of claim 21 , the first and second and third deep learning models comprising artificial convolutional neural networks or layers of an artificial convolutional neural network.

30. The system of claim 21 , wherein the instructions, when executed by one or more processors, cause the one or more processors to perform operations comprising receiving the electronic message from a message transfer application on the network.