Federated Data Influence Scoring for Malicious Instance Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional security solutions for protecting machine-learning models are ineffective in decentralized or federated settings due to the inability to inspect data from multiple sources, making them vulnerable to attacks like poisoned data and noisy training data.
Innovation Solution
The method involves calculating influence scores for data instances, ranking them based on these scores, determining anomaly scores, selecting potentially malicious data, and performing security actions to protect against it, using a system with modules for calculation, ranking, determining, selection, and security actions on federated client computing devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional security solutions are used in centralized settings, then data inspection capability is available, but these solutions are ineffective in decentralized federated settings
Solution Approach 1:
The patent segments the centralized security inspection function into distributed components that operate at each federated client device. Instead of a single centralized inspection point, each client independently calculates influence scores and performs anomaly detection on its own data instances, enabling security functionality to work in decentralized federated settings while maintaining effectiveness.
2Reliability
If all data instances are processed for security verification, then comprehensive security coverage is achieved, but computational complexity increases
Solution Approach 1:
The patent applies local quality by making each data instance's security verification tailored to its specific characteristics. The influence score calculation customizes the impact assessment for each individual data instance based on its unique properties and relationships with training data parameters, rather than applying a uniform verification process to all data. This enables comprehensive security coverage while optimizing computational resources for each specific case.
Solution Approach 2:
The patent implements partial action by focusing security verification efforts on data instances that are most likely to be malicious. By calculating influence scores and ranking data instances, the system identifies and prioritizes verification of high-risk instances, rather than expending equal computational resources on all data instances. This approach achieves effective security coverage with reduced overall computational complexity.
3Measurement precision
If influence scores are calculated for all data instances, then malicious data identification accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing training data parameters before the actual security verification process. The system prepares influence function components and relationships in advance, so that when data instances need to be evaluated, the computational work has already been partially completed. This preliminary preparation maintains high identification accuracy while reducing the time required for actual malicious data detection.
Solution Approach 2:
The patent utilizes parameter changes by transforming the security verification problem into a parameter-based influence score calculation. Instead of directly analyzing data content for maliciousness, the system changes the verification parameters to measure the influence of each data instance on training data parameters. This parameter transformation enables accurate malicious data identification through mathematical computation rather than time-consuming content analysis.
Data Source
AI summary
The disclosed computer-implemented method for verifying decentralized federated data using influence evaluation may include (i) calculate an influence score for each of a group of data instances, (ii) rank the data instances based on the influence scores, (iii) determine an anomaly score for each of the ranked data instances, (iv) select the ranked data instances with the highest anomaly scores as containing potentially malicious data, and (v) perform a security action that protects against the potentially malicious data. Various other methods, systems, and computer-readable media are also disclosed.