Adaptive detection of security threats through retraining of computer-implemented models
Adapting detection of security threats, including by retraining computer-implemented models is disclosed. An indication is received that a natural language processing model should be retrained. A list of training samples is generated that includes at least one synthetic training sample. The natural language processing model is retrained at least in part by using the set of generated training samples. The retrained natural language processing model is used to determine a likelihood that a message poses a risk.
1. A system, comprising:
a processor configured to:
receive an indication that an existing natural language processing model, trained to classify textual content, should be retrained;
generate a set comprising a plurality of training samples, wherein the set includes at least one synthetic training sample constructed using one or more linguistic hints, comprising at least one keyword of phrase, about a particular novel attack for which malicious textual communications associated with the novel attack, when processed by the existing natural language processing model, are erroneously classified as benign textual communications;
retrain the existing natural language processing model at least in part by using the set of generated training samples; and
use the retrained model to determine a likelihood that a received communication transmitted by a sender to a recipient poses a risk; and
a memory coupled to the processor and configured to provide the processor with instructions.
2. The system of claim 1 , wherein the indication is received in response to a report, by a recipient, of a malicious email.
3. The system of claim 1 , wherein the indication is received in response to a specification of the novel attack.
4. The system of claim 1 , wherein the indication includes an indication that a threshold quantity of false negatives is available for training purposes.
5. The system of claim 1 , wherein generating the set of training samples includes performing a text augmentation on an exemplar malicious message.
6. The system of claim 5 , wherein performing the text augmentation includes programmatically altering individual characters of the exemplar malicious message.
7. The system of claim 5 , wherein performing the text augmentation includes programmatically altering one or more words of the exemplar malicious message.
8. The system of claim 5 , wherein performing the text augmentation includes programmatically altering one or more clauses of the exemplar malicious message.
9. The system of claim 5 , wherein performing the text augmentation includes replacing an element of the exemplar malicious message with a semantic equivalent.
10. The system of claim 9 , wherein the semantic equivalent comprises a word synonym.
11. The system of claim 9 , wherein the semantic equivalent comprises a phrase.
12. The system of claim 1 , wherein generating the set of training samples includes modifying a predetermined amount of an exemplar malicious message.
13. The system of claim 1 , wherein retraining the natural language processing model includes training a word-level convolutional neural network that predicts an attack label of an email.
14. The system of claim 1 , wherein retraining the natural language processing model includes training a character-level convolutional neural network that predicts an attack label of an email.
15. The system of claim 1 , further comprising taking a remedial action in response to determining that the received communication poses the risk.
16. The system of claim 1 , wherein at least one hint is a reported false negative message.
17. The system of claim 1 , wherein at least one hint is a sample attack message.
18. A method, comprising:
receiving an indication that an existing natural language processing model, trained to classify textual content, should be retrained;
generating a set comprising a plurality of training samples, wherein the set includes at least one synthetic training sample constructed using one or more linguistic hints, comprising at least one keyword or phrase, about a particular novel attack for which malicious textual communications associated with the novel attack, when processed by the existing natural language processing model, are erroneously classified as benign textual communications;
retraining the existing natural language processing model at least in part by using the set of generated training samples; and
using the retrained model to determine a likelihood that a received communication transmitted by a sender to a recipient poses a risk.
19. The method of claim 18 , wherein the indication is received in response to a report, by a recipient, of a malicious email.
20. The method of claim 18 , wherein the indication is received in response to a specification of the novel attack.
21. The method of claim 18 , wherein the indication includes an indication that a threshold quantity of false negatives is available for training purposes.
22. The method of claim 18 , wherein generating the set of training samples includes performing a text augmentation on an exemplar malicious message.
23. The method of claim 22 , wherein performing the text augmentation includes programmatically altering individual characters of the exemplar malicious message.
24. The method of claim 22 , wherein performing the text augmentation includes programmatically altering one or more words of the exemplar malicious message.
25. The method of claim 22 , wherein performing the text augmentation includes programmatically altering one or more clauses of the exemplar malicious message.
26. The method of claim 22 , wherein performing the text augmentation includes replacing an element of the exemplar malicious message with a semantic equivalent.
27. The method of claim 26 , wherein the semantic equivalent comprises a word synonym.
28. The method of claim 26 , wherein the semantic equivalent comprises a phrase.
29. The method of claim 18 , wherein generating the set of training samples includes modifying a predetermined amount of an exemplar malicious message.
30. The method of claim 18 , wherein retraining the natural language processing model includes training a word-level convolutional neural network that predicts an attack label of an email.
31. The method of claim 18 , wherein retraining the natural language processing model includes training a character-level convolutional neural network that predicts an attack label of an email.
32. The method of claim 18 , further comprising taking a remedial action in response to determining that the received communication poses the risk.
33. The method of claim 18 , wherein at least one hint is a reported false negative message.
34. The method of claim 18 , wherein at least one hint is a sample attack message.