Techniques to detect perturbation attacks with an actor-critic framework
Embodiments discussed herein may be generally directed to systems and techniques to generate a quality score based on an observation and an action caused by an actor agent during a testing phase. Embodiments also include determining a temporal difference between the quality score and a previous quality score based on a previous observation and a previous action, determining whether the temporal difference exceeds a threshold value, and generating an attack indication in response to determining the temporal difference exceeds the threshold value.
1. An apparatus, comprising:
memory to store instructions; and
processing circuitry coupled with the memory:
an actor agent, executable by the processing circuitry, to cause an action in a processing environment based on an observation during a testing phase;
a critic agent, executable by the processing circuitry, to:
generate a quality score based on the observation and the action caused by the actor agent during the testing phase; and
a temporal difference detector, executable by the processing circuitry, to:
determine a temporal difference between the quality score and a previous quality score based on a previous observation and a previous action,
determine whether the temporal difference exceeds a threshold value,
generate an attack indication in response to determining the temporal difference exceeds the threshold value, and
permit processing of a next observation and a next action in response to determining the temporal difference does not exceed the threshold value.
2. The apparatus of claim 1 , wherein the attack indication to indicate an occurrence of an attack via an input in the processing environment, the attack comprising one or more of a Fast Gradient Sign Method (FGSM) attack and a random attack.
3. The apparatus of claim 1 , the actor agent to cause a series of actions including the action and the previous action, and the critic agent to determine a sequence of quality scores based on each action of the series of actions and an associated observation for each action.
4. The apparatus of claim 3 , the critic agent to determine temporal differences between quality scores of consecutive actions of the series of actions and the associated observations.
5. The apparatus of claim 4 , the critic agent to:
determine whether each of the temporal differences exceeds the threshold value;
permit the actor agent to cause a next action in the processing environment in response to determining a temporal difference of the temporal differences does not exceed the threshold value; and
generate an attack indication and prohibit the actor agent from causing a next action, in response to determining a temporal difference of the temporal differences exceeds the threshold value.
6. The apparatus of claim 1 , wherein the critic agent to perform a training phase prior to generating the quality score and generating the quality score during the testing phase.
7. The apparatus of claim 6 , the critic agent to train the actor agent during the training phase using temporal difference learning.
8. The apparatus of claim 1 , comprising one or more sensor devices to generate data for the processing environment, the sensor devices comprising at least one of a camera, a laser range finder, a radio detection and ranging (RADAR) device, a global positioning system (GPS) device, an ultrasonic device, a sound detection and ranging (SONAR) device, an altimeter, a gyroscope, a tachymeter, and an accelerometer.
9. The apparatus of claim 1 , comprising:
a storage to store a sequence of quality scores including the quality score, each quality score of the sequence of quality scores utilized to determine a temporal difference.
10. A computer-implemented method, comprising:
generating a quality score based on an observation and an action caused by an actor agent during a testing phase, the action performed in a processing environment based on the observation;
determining a temporal difference between the quality score and a previous quality score based on a previous observation and a previous action;
determining whether the temporal difference exceeds a threshold value;
generating an attack indication in response to determining the temporal difference exceeds the threshold value; and
permitting processing of a next observation and a next action in response to determining the temporal difference does not exceed the threshold value.
11. The computer-implemented method of claim 10 , wherein the attack indication indicates an occurrence of an attack via an input in the processing environment, the attack comprising one or more of a Fast Gradient Sign Method (FGSM) attack and a random attack.
12. The computer-implemented method of claim 10 , comprising:
causing a series of actions including the action and the previous action; and
determining a sequence of quality scores based on each action of the series of actions and an associated observation for each action.
13. The computer-implemented method of claim 12 , comprising determining temporal differences between quality scores of consecutive actions of the series of actions and the associated observations.
14. The computer-implemented method of claim 13 , comprising:
determining whether each of the temporal differences exceeds the threshold value;
permitting a next action in the processing environment in response to determining a temporal difference of the temporal differences does not exceed the threshold value; and
generating an attack indication and prohibiting a next action, in response to determining a temporal difference of the temporal differences exceeds the threshold value.
15. The computer-implemented method of claim 10 , comprising performing a training phase prior to generating the quality score and generating the quality score during the testing phase.
16. The computer-implemented method of claim 15 , comprising performing the training during the training phase using temporal difference learning.
17. The computer-implemented method of claim 10 , comprising receiving data from one or more sensor devices in the processing environment, the sensor devices comprising at least one of a camera, a laser range finder, a radio detection and ranging (RADAR) device, a global positioning system (GPS) device, an ultrasonic device, a sound detection and ranging (SONAR) device, an altimeter, a gyroscope, a tachymeter, and an accelerometer.
18. A non-transitory machine-readable medium containing instructions, which when executed by a processor, cause the processor to perform operations, the operations to:
generate a quality score based on an observation and an action caused by an actor agent during a testing phase, the action performed in a processing environment based on the observation;
determine a temporal difference between the quality score and a previous quality score based on a previous observation and a previous action;
determine whether the temporal difference exceeds a threshold value;
generate an attack indication in response to determining the temporal difference exceeds the threshold value; and
permit processing of a next observation and a next action in response to determining the temporal difference does not exceed the threshold value.
19. The machine-readable medium of claim 18 , wherein the attack indication indicates an occurrence of an attack via an input in the processing environment, the attack comprising one or more of a Fast Gradient Sign Method (FGSM) attack and a random attack.
20. The machine-readable medium of claim 18 , wherein the operations further comprise operations to:
cause a series of actions including the action and the previous action; and
determine a sequence of quality scores based on each action of the series of actions and an associated observation for each action.
21. The machine-readable medium of claim 20 , wherein the operations further comprise operations to determine temporal differences between quality scores of consecutive actions of the series of actions and the associated observations.
22. The machine-readable medium of claim 21 , wherein the operations further comprise operations to:
determine whether each of the temporal differences exceeds the threshold value;
permit a next action in the processing environment in response to determining a temporal difference of the temporal differences does not exceed the threshold value; and
generate an attack indication and prohibiting a next action, in response to determining a temporal difference of the temporal differences exceeds the threshold value.
23. The machine-readable medium of claim 18 , wherein the operations further comprise operations to perform a training phase prior to generating the quality score and generating the quality score during the testing phase.
24. The machine-readable medium of claim 23 , wherein the operations further comprise operations to perform the training during the training phase using temporal difference learning.
25. The machine-readable medium of claim 18 , wherein the operations further comprise operations to receive data from one or more sensor devices in the processing environment, the sensor devices comprising at least one of a camera, a laser range finder, a radio detection and ranging (RADAR) device, a global positioning system (GPS) device, an ultrasonic device, a sound detection and ranging (SONAR) device, an altimeter, a gyroscope, a tachymeter, and an accelerometer.