|
Work item:
|
X.srm-fml
|
|
Subject/title:
|
Security requirements and measures of federated machine learning
|
|
Status:
|
[Carried to next study period]
|
|
Approval process:
|
TAP
|
|
Type of work item:
|
Recommendation
|
|
Version:
|
New
|
|
Equivalent number:
|
-
|
|
Timing:
|
-
|
|
Liaison:
|
ITU-T SG2, ITU-T SG16, ITU-T SG20, IEEE C/AISC/SPFML, IEEE C/AISC/FML
|
|
Supporting members:
|
China Information Communication Technologies Group, China Unicom, China Telecom
|
|
Summary:
|
Federated machine learning (FML) is a framework or system that enables multiple participants to collaboratively build and use machine learning models without disclosing the raw and private data owned by the participants while achieving good performance. In other words, original data can not be transmitted to other clients or server during the whole process of federated machine learning; data usage occurs where the original data is. FML effectively solves the problem of data silos and achieves data sharing while ensuring privacy.
The typical stages of FML include data preparation, model training and model deployment. Although FML can solve the privacy problem of local data to a certain extent, it brings new exploitable opportunities for attackers when model parameters are shared and models are aggregated. Since FML emphasizes on distributed computing between different entities, there are many threats specific to FML compared with traditional machine learning, such as free riding attacks and coordinator vulnerability in the data preparation and model training stage, eavesdropping and model poisoning attacks in the modeling training stage, and membership inference attack in the model deployment stage etc. There are many security risks and threats to be considered in different stage of FML. These security risks and threats can seriously affect the functional correctness and data security of FML, restricting its application in various industries. In the same time, FML needs to use the corresponding security measures in advance to mitigate or defend against those security risks and threats.
Therefore, to ensure the security and data protection, it is valuable and important to study security requirements and measures of FML.
This draft new Recommendation analyzes security risks and threats of federated machine learning in the life cycle stages and provides security requirements and security measures that could mitigate or defend against the security risks and threats.
|
|
Comment:
|
incubation queue
|
|
Reference(s):
|
|
|
Historic references:
|
|
Contact(s):
|
|
| ITU-T A.5 justification(s): |
|
|
|
|
First registration in the WP:
2024-03-11 12:37:12
|
|
Last update:
2024-09-03 10:48:35
|
