Share-Based Homomorphic Encryption for Federated Gradient Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning systems face challenges in maintaining confidentiality of gradient information and incurring high communication costs, particularly when using a trusted third party (TTP) for key management, and require all edge nodes to be available for training, leading to process interruptions if some nodes become incommunicable.
Innovation Solution
A federated learning system utilizing homomorphic encryption and secret sharing to securely aggregate gradient information without a TTP, where edge devices generate encrypted gradients using a common public key, and the server aggregates and decrypts using edge-specific shares, maintaining confidentiality and reducing communication costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a trusted third party (TTP) is used to manage encryption keys in federated learning, then security and confidentiality of gradient information is improved, but communication costs increase and system complexity increases
Solution Approach 1:
The patent extracts the TTP from the system by implementing a decentralized key management mechanism where each edge device generates and manages its own key pair. The server cannot access individual device keys, eliminating the need for a separate TTP entity while maintaining security.
Solution Approach 2:
The patent segments the key management function by giving each edge device its own independent key pair instead of using a centralized TTP. This segmentation distributes trust across multiple independent entities, reducing system complexity while maintaining confidentiality.
2Reliability
If all edge nodes are required to be available during AI model training, then security is maintained, but system reliability deteriorates when nodes become incommunicable
Solution Approach 1:
The patent implements partial action by allowing the training process to continue with a subset of available edge nodes rather than requiring all nodes. The federated learning can proceed with whatever nodes are communicable, and the model will still improve, albeit at a potentially reduced rate.
Solution Approach 2:
The system dynamically adapts to the availability of edge nodes during training. The server can adjust which nodes participate in each training round based on their communicability, allowing the system to remain flexible and productive even when node availability changes.
3Reliability
If homomorphic encryption is used to protect gradient information, then confidentiality is improved, but communication volume increases
Solution Approach 1:
The patent changes the encryption parameter by using standard public key encryption for gradient transmission instead of full homomorphic encryption. This parameter change reduces communication overhead while maintaining adequate confidentiality through the combination of encryption and secure aggregation protocols.
Data Source
AI summary
It is provided a federated learning system for aggregating gradient information representing a result of training an AI model in an edge device, the federated learning system comprising the edge device and a server apparatus, the training module in the edge device being configured to generate an edge switch share in which the encrypted aggregated gradient is encrypted, and to transmit the generated edge switch share to the server apparatus, the encryption/decryption module in the server apparatus being configured to generate an encrypted aggregated gradient for decryption by adding edge switch shares received from the plurality of the edge device, generate an aggregated gradient by decrypting the generated encrypted aggregated gradient for decryption, and to transmit the generated aggregated gradient to the edge device, the training module in the edge device being configured to train the AI model by using the aggregated gradient received from the server apparatus.