Incremental updates to malware detection models
A machine learning model is sequentially fine-tuned with new training data as the training data becomes available. By using a suitable mix of old and new data, and weighting samples in the training data by age, the model can be efficiently updated to maintain accuracy against a changing malware landscape without manual modifications of the network layers or the computational expense of full retraining.
1. A computer program product for updating a previously-trained machine learning malware classification model with new samples, the computer program product comprising computer executable code embodied in a non-transitory computer-readable medium that, when executing on one or more computing devices, performs the steps of:
receiving new samples to be used to update a previously-trained machine learning malware classification model;
determining a number of the new samples to be selected for use in updating;
determining a number of past samples to be selected as a percentage of the number of the new samples to be selected for updating;
selecting new samples based on the determined number of new samples to provide selected new samples;
randomly selecting from past samples based on the determined number of past samples to provide a predetermined number of randomly selected past samples; and
updating the previously-trained model using the selected new samples and the predetermined number of randomly selected past samples.
2. The computer program product of claim 1 , wherein the percentage is a percentage is in a range from 80% to 120%.
3. A method for updating a trained model with new samples, the method comprising:
receiving new samples to be used to update a model;
determining a number of the new samples to be used for updating;
determining a number of past samples to select based on the number of the new samples to be used for updating;
selecting new samples based on the determined number of new samples to provide selected new samples;
randomly selecting from past samples based on the determined number of past samples to provide a predetermined number of randomly selected past samples; and
updating the trained model using the selected new samples and the predetermined number of randomly selected past samples.
4. The method of claim 3 , wherein updating the trained model comprises using machine learning techniques to train an existing model using the selected new samples and the predetermined number of randomly selected past samples.
5. The method of claim 3 , wherein the model is a malware classifier.
6. The method of claim 3 , where in the model is a maliciousness classifier.
7. The method of claim 3 , wherein the number of past samples is determined as a percentage or a multiple of the number of new samples to be used for updating.
8. The method of claim 3 , wherein the number of past samples is determined based on a desired ratio of past samples and new samples.
9. The method of claim 3 , wherein the method is performed periodically to update the trained model using recently collected samples as new samples.
10. The method of claim 3 , wherein the number of new samples is determined by identifying a number of new samples to be used as test data.
11. The method of claim 3 , wherein the number of new samples is determined by identifying a number of new samples to be used as training data.
12. The method of claim 3 , wherein the predetermined number of randomly selected past samples are randomly selected with weighting toward more recent past samples.
13. The method of claim 3 , wherein the predetermined number of randomly selected past samples are randomly selected with weighting toward a statistical distribution of features in underlying data.
14. A system for maintaining a trained model with new samples, the system comprising:
a sample repository including a memory for receiving new samples to be used to update the model;
a new sample number determination system for determining a number of the new samples to be used for updating and for determining a number of past samples to select based on the number of the new samples to be used for updating;
a sample selector for selecting new samples based on the determined number of new samples to provide selected new samples and for randomly selecting from past samples based on the determined number of past samples to provide a predetermined number of randomly selected past samples;
a training engine for updating the trained model using the selected new samples and the predetermined number of randomly selected past samples; and
a security agent for using the updated trained model to detect malware, wherein the new sample number determination system, the sample selector, the training engine, and the security agent include one or more processors configured by non-transitory computer executable code to perform corresponding functions.
15. The system of claim 14 , wherein the new samples are updated periodically.
16. The system of claim 14 , wherein the trained model is a malware classifier.
17. The system of claim 14 , wherein the trained model is a deep learning neural network.
18. The system of claim 14 , wherein the new sample number determination system determines the number of new samples by identifying a number of new samples to be used as training data.
19. The system of claim 14 , wherein the new sample number determination system determines the number of past samples based on a percentage of the number of new samples.
20. The system of claim 14 , wherein the predetermined number of randomly selected past samples are selected randomly with weighting toward more recent samples.