Securing a network device by forecasting an attack event using a recurrent neural network
Securing a network device by forecasting an attack event using a recurrent neural network. In one embodiment, a method may include collecting event sequences of events that occurred on multiple network devices, generating training sequences, validation sequences, and test sequences from the event sequences, training a recurrent neural network using the training sequences, the validation sequences, and the test sequences, collecting an event sequence of the most recent events that occurred on a target network device, forecasting, using the recurrent neural network and based on the event sequence of the most recent events that occurred on the target network device, the next event that will occur on the target network device, and in response to the forecasted next event being an attack event, performing a security action to prevent harm to the target network device from the attack event.
1. A computer-implemented method for securing a network device by forecasting an attack event using a recurrent neural network of a security server, at least a portion of the method being performed by a computing device comprising at least one processor, the method comprising:
collecting, by the security server, event sequences of events that occurred on multiple network devices, the events in the event sequences including benign noise events and attack events;
generating, by the security server, training sequences, validation sequences, and test sequences from the event sequences;
training, by the security server, the recurrent neural network of the security server using the training sequences, the validation sequences, and the test sequences, the recurrent neural network being capable of capturing an implicit dependency among the events;
collecting an event sequence of the most recent events that occurred on a target network device;
automatically forecasting, using the recurrent neural network and based on the event sequence of the most recent events that occurred on the target network device, the very next event that will occur in the future in the event sequence on the target network device, regardless of an event type of whether the forecasted very next event is forecasted to be a future attack event or a future benign noise event; and
in response to the forecasted very next event that will occur in the future being forecasted to be a future attack event, performing a security action to prevent harm to the target network device from the future attack event.
2. The method of claim 1 , wherein the recurrent neural network is a Long Short-Term Memory (LSTM) Memory Array (MA) recurrent neural network.
3. The method of claim 1 , wherein:
the method further comprises empirically identifying a fixed input sequence length;
the length of each of the training sequences, the validation sequences, and the test sequences is equal to the fixed input sequence length; and
the generating of the training sequences, the validation sequences, and the test sequences from the event sequences includes splitting event sequences that are longer than the fixed input sequence length into multiple training sequences, validation sequences, or test sequences, or some combination thereof.
4. The method of claim 3 , further comprising periodically updating the fixed input sequence length.
5. The method of claim 1 , wherein the event sequence of the most recent events that occurred on the target network device has a variable length.
6. The method of claim 1 , wherein each of the training sequences, the validation sequences, and the test sequences includes a label that corresponds to the next event in the corresponding sequence.
7. The method of claim 1 , wherein the collecting of the event sequences of events that occurred on the multiple network devices is performed periodically at a regular interval.
8. The method of claim 7 , wherein the regular interval is daily.
9. The method of claim 1 , wherein the forecasting includes the recurrent neural network distinguishing between attack events and benign noise events in the event sequence of the most recent events that occurred on the target network device.
10. The method of claim 9 , wherein the forecasting includes ignoring the noise events and using the attack events to forecast the very next event that will occur in the future in the event sequence on the target network device.
11. One or more non-transitory computer-readable media comprising one or more computer-readable instructions that, when executed by one or more processors of one or more computing devices, cause the one or more computing devices to perform a method for securing a network device by forecasting an attack event using a recurrent neural network of a security server, the method comprising:
collecting, by the security server, event sequences of events that occurred on multiple network devices, the events in the event sequences including benign noise events and attack events;
generating, by the security server, training sequences, validation sequences, and test sequences from the event sequences;
training, by the security server, the recurrent neural network of the security server using the training sequences, the validation sequences, and the test sequences, the recurrent neural network being capable of capturing an implicit dependency among the events;
collecting an event sequence of the most recent events that occurred on a target network device;
automatically forecasting, using the recurrent neural network and based on the event sequence of the most recent events that occurred on the target network device, the very next event that will occur in the future in the event sequence on the target network device, regardless of an event type of whether the forecasted very next event is forecasted to be a future attack event or a future benign noise event; and
in response to the forecasted very next event that will occur in the future being forecasted to be a future attack event, performing a security action to prevent harm to the target network device from the future attack event.
12. The one or more non-transitory computer-readable media of claim 11 , wherein the recurrent neural network is a Long Short-Term Memory (LSTM) Memory Array (MA) recurrent neural network.
13. The one or more non-transitory computer-readable media of claim 11 , wherein:
the method further comprises empirically identifying a fixed input sequence length; and
the length of each of the training sequences, the validation sequences, and the test sequences is equal to the fixed input sequence length; and
the generating of the training sequences, the validation sequences, and the test sequences from the event sequences includes splitting event sequences that are longer than the fixed input sequence length into multiple training sequences, validation sequences, or test sequences, or some combination thereof.
14. The one or more non-transitory computer-readable media of claim 13 , further comprising periodically updating the fixed input sequence length.
15. The one or more non-transitory computer-readable media of claim 11 , wherein the event sequence of the most recent events that occurred on the target network device has a variable length.
16. The one or more non-transitory computer-readable media of claim 11 , wherein each of the training sequences, the validation sequences, and the test sequences includes a label that corresponds to the next event in the corresponding sequence.
17. The one or more non-transitory computer-readable media of claim 11 , wherein the collecting of the event sequences of events that occurred on the multiple network devices is performed periodically at a regular interval.
18. The one or more non-transitory computer-readable media of claim 17 , wherein the regular interval is daily.
19. The one or more non-transitory computer-readable media of claim 11 , wherein the forecasting includes the recurrent neural network distinguishing between attack events and benign noise events in the event sequence of the most recent events that occurred on the target network device.
20. The one or more non-transitory computer-readable media of claim 19 , wherein the forecasting includes ignoring the noise events and using the attack events to forecast the very next event that will occur in the future in the event sequence on the target network device.