Security system using a convolutional neural network with pruned filters
Security systems and methods for detecting intrusion events include one or more sensors configured to monitor an environment. A pruned convolutional neural network (CNN) is configured process information from the one or more sensors to classify events in the monitored environment. CNN filters having the smallest summed weights have been pruned from the pruned CNN. An alert module is configured to detect an intrusion event in the monitored environment based on event classifications. A control module is configured to perform a security action based on the detection of an intrusion event.
1. A security system, comprising:
one or more sensors configured to monitor an environment;
a pruned convolutional neural network (CNN) configured process information from the one or more sensors to classify events in the monitored environment, wherein CNN filters having the smallest summed weights have been pruned from the pruned CNN; and
a control module configured to perform a security action based on event classifications.
2. The security system of claim 1 , further comprising a pruning module configured to calculate a sum of weights for each filter in a layer of a CNN, to sort the filters in the layer by respective sums of weights, to prune m filters with the smallest sums of weights to decrease a computational cost of operating the CNN.
3. The security system of claim 2 , further comprising a training module configured to retrain the pruned CNN to repair accuracy loss that results from pruning the filters.
4. The security system of claim 3 , wherein the pruning module and the training module are further configured to iterate pruning and retraining until a threshold CNN accuracy is reached.
5. The security system of claim 2 , wherein the pruning module is further configured to prune feature maps corresponding to the m pruned filters.
6. The security system of claim 5 , wherein the pruning module is further configured to prune kernels in a subsequent layer that correspond to the pruned feature maps.
7. The security system of claim 2 , wherein the pruning module is further configured to select a number of filters m based on a sensitivity of the layer to pruning.
8. The security system of claim 7 , wherein a smaller m is selected for layers that have relatively high sensitivities compared to layers that have relatively low sensitivities.
9. The security system of claim 7 , wherein sensitivity to pruning is measured as a degree of accuracy change.
10. The security system of claim 2 , wherein the training module is further configured to create a new kernel matrix for the layer.
11. The security system of claim 10 , wherein the training module is further configured to copy copying un-pruned kernel weights to the new kernel matrix.
12. A method of detecting intrusion events, comprising:
monitoring an environment using one or more sensors;
classifying events in the monitored environment based on information from the one or more sensors using a pruned convolutional neural network (CNN), wherein CNN filters having the smallest summed weights have been pruned from the pruned CNN; and
performing a security action based on event classifications.
13. The method of claim 12 , further comprising pruning the CNN by calculating a sum of weights for each filter in a layer of the CNN, sorting the filters in the layer by respective sums of weights, pruning m filters with the smallest sums of weights to decrease a computational cost of operating the CNN, and retraining the pruned CNN to repair accuracy loss that results from pruning the filters.
14. The method of claim 13 , further comprising pruning feature maps corresponding to the m pruned filters.
15. The method of claim 14 , pruning kernels in a subsequent layer that correspond to the pruned feature maps.
16. The method of claim 13 , further comprising iterating the steps of pruning and retraining until a threshold CNN accuracy is reached.
17. The method of claim 13 , further comprising selecting a number of filters m based on a sensitivity of the layer to pruning.
18. The method of claim 17 , wherein a smaller m is selected for layers that have relatively high sensitivities compared to layers that have relatively low sensitivities.
19. The method of claim 13 , wherein retraining the CNN comprises creating a new kernel matrix for the layer.
20. The method of claim 12 , wherein the security action comprises one or more actions selected from the group consisting of image recognition, video recognition, face recognition, and natural language processing.