Decision Tree Outlier Detection Against Attribute Inference Attacks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In federated learning, malicious attackers can manipulate training data to infer attribute values, and existing systems lack protection against such attacks.
Innovation Solution
An information processing apparatus that detects outliers in decision tree outputs and determines vulnerability to attribute inference attacks, and generates a preliminarily-manipulated decision tree with candidate nodes to facilitate inference of attribute values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning is implemented to train machine learning models across multiple clients, then model training capability and data privacy are improved, but vulnerability to attribute inference attacks increases due to malicious clients performing preliminary manipulation
Solution Approach 1:
The patent applies preliminary action by detecting outliers in decision tree output values before they can be exploited for attribute inference attacks. The system proactively identifies malicious manipulation patterns in the federated learning process, allowing preventive measures to be taken before the attack can successfully infer sensitive attribute values from training data.
Solution Approach 2:
The patent introduces an intermediary detection mechanism that monitors the decision tree output values during federated learning. This intermediary system acts as a mediator between the training process and potential attacks, identifying abnormal patterns that indicate malicious clients attempting preliminary manipulation, thereby protecting the overall system without stopping the federated learning process.
2Productivity
If decision trees are used in federated learning for their interpretability and efficiency, then training speed and model interpretability are improved, but susceptibility to outlier-based attacks increases
Solution Approach 1:
The patent implements feedback by continuously monitoring decision tree output values during the federated learning process. When outliers are detected, the system provides feedback to identify potential malicious clients, allowing the system to respond to attacks while maintaining the efficient decision tree structure and training speed that make federated learning productive.
Data Source
AI summary
Provided is a technique to prevent an attribute inference attack against a federated learning model. An information processing apparatus (2) includes: a detection section (21) that detects, as an outlier, an output value which falls outside a normal range in terms of magnitude of a value among output values of a decision tree which is undergoing federated learning; and a determination section (22) that determines, in a case where the outlier has been detected, that the decision tree has vulnerability to an attack to infer an attribute value included in training data which has been used to train the decision tree.