Machine learning adversarial campaign mitigation on a computing device
Machine learning adversarial campaign mitigation on a computing device. The method may include deploying an original machine learning model in a model environment associated with a client device; deploying a classification monitor in the model environment to monitor classification decision outputs in the machine learning model; detecting, by the classification monitor, a campaign of adversarial classification decision outputs in the machine learning model; applying a transformation function to the machine learning model in the model environment to transform the adversarial classification decision outputs to thwart the campaign of adversarial classification decision outputs; determining a malicious attack on the client device based in part on detecting the campaign of adversarial classification decision outputs; and implementing a security action to protect the computing device against the malicious attack.
1. A computer-implemented method for mitigation of machine learning adversarial campaigns on a computing device, at least a portion of the method being performed by the computing device comprising one or more processors, the method comprising:
deploying a machine learning model in a model environment associated with the computing device;
deploying a classification monitor in the model environment to monitor classification decision outputs in the machine learning model, wherein the deploying of the classification monitor in the model environment comprises deploying the classification monitor in a threat environment where greater than 95% of the classification decision outputs have an equal to or less than 0.1% false positive rate;
detecting, by the classification monitor, a campaign of adversarial classification decision outputs in the machine learning model;
applying a transformation function to the machine learning model in the model environment to transform the adversarial classification decision outputs to thwart the campaign of adversarial classification decision outputs;
determining a malicious attack on the computing device based in part on detecting the campaign of adversarial classification decision outputs; and
implementing a security action to protect the computing device against the malicious attack.
2. The method of claim 1 , wherein the deploying of the machine learning model further comprises:
deploying a binary classifier.
3. The method of claim 1 , further comprising:
determining a cessation of the campaign of adversarial classification decision outputs;
and restoring the machine learning model in the model environment after the cessation.
4. The method of claim 1 , wherein the monitoring of the classification decision outputs further comprises monitoring a scalar score.
5. The method of claim 1 , wherein the monitoring of the classification decision outputs further comprises:
determining that each of the classification decision outputs is either malicious or benign.
6. The method of claim 1 , wherein the detecting of the campaign of adversarial classification decision outputs further comprises:
detecting a plurality of adversarial inputs during runtime on the computing device that are intended to misclassify the classification decision outputs.
7. The method of claim 6 , further comprising:
evaluating a performance level of the machine learning model by analyzing a receiver operating characteristics (ROC) curve.
8. The method of claim 7 , further comprising:
setting an ROC threshold, the threshold associated with a pre-determined point on the ROC curve;
deriving a binary difficulty detector from a distribution of naturally-occurring difficult samples;
sampling from the distribution at different proportions of binary ground-truth labels;
proactively computing a low-bandwidth equiprobable posterior function from the sampling based on detecting the plurality of adversarial inputs during runtime; and
deploying the low-bandwidth equiprobable posterior function based on each of a plurality of adversarial environments.
9. The method of claim 7 , further comprising:
detecting a pinch-down on the ROC curve; and
applying the transformation function in the model environment to transform the adversarial classification decision outputs based on detecting the pinch-down on the ROC curve.
10. The method of claim 1 , wherein the applying of the transformation function in the model environment further comprises:
deriving a low-bandwidth equiprobable posterior function.
11. The method of claim 1 , wherein the applying of the transformation function in the model environment further comprises:
determining a set of classification decision outputs that satisfy a pre-determined plurality of scalar score intervals.
12. A system for machine learning adversarial campaign mitigation on a computing device, comprising:
a processor;
a memory in electronic communication with the processor; and
instructions stored in the memory, the instructions being executable by the processor to:
deploy a machine learning model in a model environment associated with the computing device;
deploy a classification monitor in the model environment to monitor classification decision outputs in the machine learning model, wherein the deploying of the classification monitor in the model environment comprises deploying the classification monitor in a threat environment where greater than 95% of the classification decision outputs have an equal to or less than 0.1% false positive rate
detect, by the classification monitor, a campaign of adversarial classification decision outputs in the machine learning model;
apply a transformation function to the machine learning model in the model environment to transform the adversarial classification decision outputs to thwart the campaign of adversarial classification decision outputs;
determine a malicious attack on the computing device based in part on detecting the campaign of adversarial classification decision outputs; and
implement a security action to protect the computing device against the malicious attack.
13. The system of claim 12 , where the instructions are further executable to:
determine a cessation of the campaign of adversarial classification decision outputs; and
restore the machine learning model in the model environment after the cessation.
14. The system of claim 12 , wherein, when the processor monitors the classification decision outputs, the instructions are further executable to monitor a scalar score.
15. The system of claim 12 , wherein, when the processor monitors the classification decision outputs, the instructions are further executable to:
determine that each of the classification decision outputs is either malicious or benign.
16. The system of claim 12 , wherein, when the processor detects the campaign of adversarial classification decision outputs, the instructions are further executable to:
detect a plurality of adversarial inputs during runtime on the computing device that are intended to misclassify the classification decision outputs.
17. The system of claim 16 , the instructions being further executable to:
evaluate a performance level of the machine learning model by analyzing a receiver operating characteristics curve (ROC).
18. The system of claim 17 , the instructions being further executable to:
set an ROC threshold, the threshold associated with a pre-determined point on the ROC curve;
derive a binary difficulty detector from a distribution of naturally-occurring difficult samples;
sample from the distribution at different proportions of binary ground-truth labels;
proactively compute a low-bandwidth equiprobable posterior function from the sampling based on detecting the plurality of adversarial inputs during runtime; and
deploy the low-bandwidth equiprobable posterior function based on each of a plurality of adversarial environments.
19. The system of claim 17 , the instructions being further executable to:
detect a pinch-down on the ROC curve; and
apply the transformation function in the model environment to transform the adversarial classification decision outputs based on detecting the pinch-down on the ROC curve.
20. One or more non-transitory computer-readable media comprising one or more computer readable instructions that, when executed by one or more processors of a computing device, cause the computing device to perform a method for machine learning adversarial campaign mitigation on a computing device, comprising:
deploying a machine learning model in a model environment associated with the computing device;
deploying a classification monitor in the model environment to monitor classification decision outputs in the machine learning model, wherein the deploying of the classification monitor in the model environment comprises deploying the classification monitor in a threat environment where greater than 95% of the classification decision outputs have an equal to or less than 0.1% false positive rate;
detecting, by the classification monitor, a campaign of adversarial classification decision outputs in the machine learning model;
applying a transformation function to the machine learning model in the model environment to transform the adversarial classification decision outputs to thwart the campaign of adversarial classification decision outputs;
determining a malicious attack on the computing device based in part on detecting the campaign of adversarial classification decision outputs; and
implementing a security action to protect the computing device against the malicious attack.