Federated Learning Authentication Embeddings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing user authentication models using federated learning are not privacy-preserving, as they require direct access to sensitive user data and embeddings, and conventional methods like one-hot encoding are not scalable and increase computational complexity, especially in resource-constrained devices.
Innovation Solution
The method generates embedding vectors using error correction codes, ensuring minimum pairwise distance between embeddings while preserving privacy, and trains neural network models using a sigmoid activation function to maximize correlation between model outputs and user embeddings, allowing for scalable and privacy-preserving user authentication.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional federated learning methods are used for user authentication, then model training can be performed distributedly, but privacy is not preserved because servers can access and store user embeddings
Solution Approach 1:
The patent introduces an intermediary mechanism where users generate their own embeddings locally using shared public parameters, eliminating the need for servers to store or access private user embeddings. The server only receives encrypted model updates, not raw data or embeddings, thus preserving privacy while maintaining distributed training capability
Solution Approach 2:
Instead of servers storing original user embeddings, the system creates encrypted copies of embedding information that can be transmitted and processed without revealing the actual embedding values. The server works with encrypted representations that preserve authentication functionality while preventing direct access to sensitive data
2Productivity
If one-hot encoding is used for user embeddings, then user identification can be performed, but computational complexity and storage requirements increase significantly
Solution Approach 1:
The patent changes the parameter representation from high-dimensional one-hot vectors to compact fixed-length embedding vectors. Instead of using n-dimensional one-hot encodings for n users, the system uses a fixed embedding dimension that is independent of the number of users, significantly reducing computational complexity and storage requirements while maintaining authentication capability
Solution Approach 2:
The system transitions from a dimensionality that scales with the number of users (one-hot encoding requires n dimensions for n users) to a fixed-dimensional embedding space. This dimensional transformation allows the system to maintain user discrimination capability while using constant computational resources regardless of user base size
3Manufacturing precision
If embeddings are generated without error correction codes, then the system is simpler, but minimum pairwise distance between embeddings cannot be guaranteed
Solution Approach 1:
The patent applies preliminary error correction code encoding to embedding vectors before they are used in authentication. This pre-processing step ensures that even if embeddings are perturbed or corrupted during transmission or storage, the minimum pairwise distance between different users' embeddings is maintained, preventing authentication failures while adding minimal computational overhead
Data Source
AI summary
Certain aspects of the present disclosure provide techniques for authenticating a user based on a machine learning model, including receiving user authentication data associated with a user; generating output from a neural network model based on the user authentication data; determining a distance between the output and an embedding vector associated with the user; comparing the determined distance to a distance threshold; and making an authentication decision based on the comparison.