Systems and methods for detecting potentially malicious content in decentralized machine-learning model updates
The disclosed computer-implemented method for detecting potentially malicious content in decentralized machine-learning model updates may include (i) receiving messages communicated within a group of client devices for performing an update of a shared machine-learning model, (ii) determining a bias of a target message in the messages communicated from a target client device in the group with respect to a remaining number of the messages in the messages communicated from the other client devices in the group, (iii) assigning a confidence score to each of the other client devices based on the bias determined for the target message, the confidence score representing a likelihood of potentially malicious content in the target message, and (iv) performing, based on the confidence score, a security action that prevents the potentially malicious content from compromising the update of the shared machine-learning model. Various other methods, systems, and computer-readable media are also disclosed.
1. A computer-implemented method for detecting potentially malicious content in decentralized machine-learning model updates, at least a portion of the method being performed by one or more computing devices comprising at least one processor, the method comprising:
receiving, by the one or more computing devices, a plurality of messages communicated within a group of client devices for performing an update of a shared machine-learning model;
determining, by the one or more computing devices, a bias of a target message in the plurality of messages communicated from a target client device in the group with respect to a remaining number of the messages in the plurality of messages communicated from the other client devices in the group;
assigning, by the one or more computing devices, a confidence score to each of the other client devices based on the bias determined for the target message, the confidence score representing a likelihood of potentially malicious content in the target message; and
performing, by the one or more computing devices and based on the confidence score, a security action that prevents the potentially malicious content from compromising the update of the shared machine-learning model.
2. The computer-implemented method of claim 1 , further comprising:
receiving a plurality of new messages for performing the update of the shared machine-learning model from additional client devices outside of the group of the client devices;
assigning a weight to the additional client devices; and
minimizing, based on the assigned weight, an influence of the new messages communicated from the additional client devices when performing the update of the shared machine-learning model.
3. The computer-implemented method of claim 2 , wherein the plurality of messages communicated within the group of client devices and the new messages communicated from the additional client devices are combined into the shared-machine learning model over a plurality of iterations.
4. The computer-implemented method of claim 2 , wherein the additional client devices have a trusted relationship level with the group of the client devices.
5. The computer-implemented method of claim 4 , wherein a pre-determined trust relationship level shared by the group of the client devices exceeds the trusted relationship level of the additional client devices.
6. The computer-implemented method of claim 1 , wherein determining the bias of the target message comprises:
calculating a ratio of a plurality of message variables associated with the messages communicated from the other client devices and a plurality of message variables associated with the target message communicated from the target client device; and
determining the bias of the target message based on a size of the calculated ratio.
7. The computer-implemented method of claim 1 , wherein assigning the confidence score comprises:
determining, based on the bias, that the target message represents a large deviation from at least one of:
the remaining number of messages; or
the shared machine-learning model; and
assigning a value to the other client devices indicating a high likelihood of the potentially malicious content in the target message.
8. The computer-implemented method of claim 1 , wherein assigning the confidence score comprises:
determining, based on the bias, that the target message represents a small deviation from at least one of:
the remaining number of messages; or
the shared-machine learning model; and
assigning a value to the other client devices indicating a low likelihood of the potentially malicious content in the target message.
9. The computer-implemented method of claim 1 , wherein assigning the confidence score comprises updating a previously determined confidence score for each of the other client devices based on the bias determined for the target message.
10. The computer-implemented method of claim 1 , wherein performing the security action comprises minimizing an impact of the potentially malicious content by filtering the potentially malicious content from the update of the shared machine-learning model.
11. The computer-implemented method of claim 1 , wherein each of the client devices in the group share at least one of a learned and a pre-defined trust relationship and wherein, based on the trust relationships, each of the client devices learns a custom machine-learning model.
12. A system for detecting potentially malicious content in decentralized machine-learning model updates, the system comprising:
at least one physical processor;
physical memory comprising a plurality of modules and computer-executable instructions that, when executed by the physical processor, cause the physical processor to:
receive, by a receiving module, a plurality of messages communicated within a group of client devices for performing an update of a shared machine-learning model;
determine, by a determining module, a bias of a target message in the plurality of messages communicated from a target client device in the group with respect to a remaining number of the messages in the plurality of messages communicated from the other client devices in the group;
assign, by an assignment module, a confidence score to each of the other client devices based on the bias determined for the target message, the confidence score representing a likelihood of potentially malicious content in the target message; and
perform, by a security module and based on the confidence score, a security action that prevents the potentially malicious content from compromising the update of the shared machine-learning model.
13. The system of claim 12 , wherein the assignment module:
receives a plurality of new messages for performing the update of the shared machine-learning model from additional client devices outside of the group of the client devices;
assigns a weight to the additional client devices; and
minimizes, based on the assigned weight, an influence of the new messages communicated from the additional client devices when performing the update of the shared machine-learning model.
14. The system of claim 13 , wherein the plurality of messages communicated within the group of client devices and the new messages communicated from the additional client devices are combined into the shared-machine learning model over a plurality of iterations.
15. The system of claim 13 , wherein the additional client devices have a trusted relationship level with the group of the client devices.
16. The system of claim 15 , wherein a pre-determined trust relationship level shared by the group of the client devices exceeds the trusted relationship level of the additional client devices.
17. The system of claim 12 , wherein the determining module determines the bias of the target message by:
calculating a ratio of a plurality of message variables associated with the messages communicated from the other client devices and a plurality of message variables associated with the target message communicated from the target client device; and
determining the bias of the target message based on a size of the calculated ratio.
18. The system of claim 12 , wherein the assignment module assigns the confidence score by:
determining, based on the bias, that the target message represents a large deviation from at least one of:
the remaining number of messages; or
the shared-machine learning model; and
assigning a value to the other client devices indicating a high likelihood of the potentially malicious content in the target message.
19. The system of claim 12 , wherein the assignment module assigns the confidence score by:
determining, based on the bias, that the target message represents a small deviation from at least one of:
the remaining number of messages; or
the shared machine-learning model; and
assigning a value to the other client devices indicating a low likelihood of the potentially malicious content in the target message.
20. A non-transitory computer-readable medium comprising one or more computer-executable instructions that, when executed by at least one processor of a computing device, cause the computing device to:
receive a plurality of messages communicated within a group of client devices for performing an update of a shared machine-learning model;
determine a bias of a target message in the plurality of messages communicated from a target client device in the group with respect to a remaining number of the messages in the plurality of messages communicated from the other client devices in the group;
assign a confidence score to each of the other client devices based on the bias determined for the target message, the confidence score representing a likelihood of potentially malicious content in the target message; and
perform, based on the confidence score, a security action that prevents the potentially malicious content from compromising the update of the shared machine-learning model.