Defending machine learning systems from adversarial attacks
Techniques are disclosed for detecting adversarial attacks. A machine learning (ML) system processes the input into and output of a ML model using an adversarial detection module that does not include a direct external interface. The adversarial detection module includes a detection model that generates a score indicative of whether the input is adversarial using, e.g., a neural fingerprinting technique or a comparison of features extracted by a surrogate ML model to an expected feature distribution for the output of the ML model. In turn, the adversarial score is compared to a predefined threshold for raising an adversarial flag. Appropriate remedial measures, such as notifying a user, may be taken when the adversarial score satisfies the threshold and raises the adversarial flag.
1. A computer-implemented method for detecting adversarial attacks on a machine-learning (ML) system, the method comprising:
receiving by an ML model of the ML system input data;
processing by the ML model the input data to generate output data;
receiving by an adversarial detection module of the ML system both the input data and the output data;
inputting a perturbed input data and the output data into a neural fingerprinting model included in the adversarial detection module, wherein the perturbed input data is generated by introducing a set of predefined random perturbations into the input data;
generating by the neural fingerprinting model a perturbed output data based on the perturbed input data;
determining using the neural fingerprinting model an adversarial score indicating whether the perturbed output data matches an expected perturbed output data for a class of data associated with the input data and the output data; and
performing one or more remedial actions based on the adversarial score.
2. The computer-implemented method of claim 1 , wherein a difference between the perturbed output data and the expected perturbed output data comprises a distance within a feature space between an output perturbation and a corresponding expected output perturbation.
3. The computer-implemented method of claim 1 , further comprising performing one or more other remedial actions when a difference between the perturbed output data and the expected perturbed output data satisfies a predefined threshold.
4. The computer-implemented method of claim 1 , further comprising:
extracting, by a surrogate ML model included in the adversarial detection module, features from the input data; and
the step of determining the adversarial score comprises comparing the features extracted by the surrogate ML model with an expected feature distribution associated with the output data.
5. The computer-implemented method of claim 4 , wherein the step of comparing the features extracted by the surrogate ML model comprises determining an energy distance between the extracted features and the expected feature distribution or a maximum mean discrepancy between the extracted features and an expected feature distribution.
6. The computer-implemented method of claim 5 , further comprising performing one or more other remedial actions when the energy distance or maximum mean discrepancy satisfies a predefined threshold.
7. The computer-implemented method of claim 1 , wherein the adversarial detection module comprises a software module capable of being implemented in multiple different ML systems without modification.
8. The computer-implemented method of claim 1 , wherein the input data includes at least one of an image, a microphone recording, a thermal camera image, LIDAR (Light Detection and Ranging) data, or RADAR data.
9. The computer-implemented method of claim 8 , wherein the ML model comprises one of a deep learning model, a support vector machine, a boosted tree, a random forest, a logistic regression model, or a linear regression model.
10. A non-transitory computer-readable storage medium including instructions that, when executed by a processor, cause the processor to perform steps for detecting adversarial attacks on a machine-learning (ML) system, the steps comprising:
receiving by an ML model of the ML system input data;
processing by the ML model the input data to generate output data;
receiving by an adversarial detection module of the ML system both the input data and the output data;
inputting a perturbed input data and the output data into a neural fingerprinting model included in the adversarial detection module, wherein the perturbed input data is generated by introducing a set of predefined random perturbations into the input data;
generating by the neural fingerprinting model a perturbed output data based on the perturbed input data;
determining using the neural fingerprinting model an adversarial score indicating whether the perturbed output data matches an expected perturbed output data for a class of data associated with the input data and the output data; and
performing one or more remedial actions based on the adversarial score.
11. The computer-readable storage medium of claim 10 , wherein the one or more remedial actions include notifying a user.
12. The computer-readable storage medium of claim 10 , wherein the one or more remedial actions include accessing an alternative source of information to classify the input data.
13. The computer-readable storage medium of claim 10 , further comprising:
extracting, by a surrogate ML model included in the adversarial detection module, features from the input data; and
the step of determining the adversarial score comprises comparing the features extracted by the surrogate ML model with an expected feature distribution associated with the output data.
14. The computer-readable storage medium of claim 13 , wherein the surrogate ML model is trained on a smaller set of training data than the ML model is trained.
15. The computer-readable storage medium of claim 13 , wherein an architecture of the surrogate ML model is less complex than an architecture of the ML model.
16. The computer-readable storage medium of claim 13 , wherein:
the step of comparing the features extracted by the surrogate ML model comprises determining an energy distance between the extracted features and the expected feature distribution or a maximum mean discrepancy between the extracted features and the expected feature distribution; and
the processor is further configured to perform one or more other remedial actions if the energy distance or maximum mean discrepancy satisfies a predefined threshold.
17. The computer-readable storage medium of claim 10 , wherein the adversarial detection module determines whether the input data is adversarial independently of a type of the ML model.
18. A system, comprising:
a memory storing a machine learning (ML) system comprising a ML model and an adversarial detection module; and
a processor that is coupled to the memory and configured to, receive by the ML model input data;
process by the ML model the input data to generate output data;
receive by the adversarial detection module both the input data and the output data;
input a perturbed input data and the output data into a neural fingerprinting model included in the adversarial detection module, wherein the perturbed input data is generated by introducing a set of predefined random perturbations into the input data;
generate by the neural fingerprinting model a perturbed output data based on the perturbed input data;
determine using the neural fingerprinting model an adversarial score indicating whether the perturbed output data matches an expected perturbed output data for a class of data associated with the input data and the output data; and
perform one or more remedial actions based on the adversarial score.
19. The computer-implemented method of claim 1 , wherein the neural fingerprinting model is different from the ML model.
20. The computer-readable storage medium of claim 10 , wherein the neural fingerprinting model is different from the ML model.