Adaptive Split Learning Face Recognition for Heterogeneous Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition technologies face challenges in addressing device heterogeneity and privacy protection in federated learning, leading to suboptimal performance and potential data leakage, especially in face recognition tasks where user devices have varying computation capabilities and data distributions.
Innovation Solution
The proposed method employs an adaptive split learning-federated learning approach, where a RingSFL system allocates propagation step lengths and aggregation weights to user terminals based on their device information, enabling efficient local joint processing and enhanced privacy protection by only transmitting output layer gradients, thereby improving training efficiency and data security.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional federated learning is used for face recognition training, then user privacy is protected to some extent, but user devices with heterogeneous computation capabilities and data distributions cause efficiency degradation and potential privacy leakage
Solution Approach 1:
The patent segments the neural network into multiple layers and assigns different segments to different user devices for training. Each device trains only its assigned segment using local data, and gradient updates are aggregated through a ring topology communication structure. This segmentation allows devices with heterogeneous capabilities to contribute according to their strengths while maintaining privacy protection.
Solution Approach 2:
The patent implements local quality by allowing each user device to perform forward propagation and backward propagation independently on its assigned network segment using its own local data and computation resources. Each device optimizes its local segment quality without exposing raw data, achieving both efficiency and privacy goals.
2Adaptability or versatility
If devices with limited computation resources participate in federated learning, then more data diversity is achieved, but these devices become stragglers causing training bias and suboptimal performance
Solution Approach 1:
By segmenting the network and assigning lighter computational segments to devices with limited resources, the system enables broader device participation while maintaining overall model performance. Each device contributes according to its capability level.
Solution Approach 2:
The system dynamically assigns network segments to devices based on their computation capabilities and data characteristics. This dynamic allocation ensures that stragglers can still contribute meaningfully without causing training bias, adapting to heterogeneous device conditions.
3Loss of information
If full model training is performed on each user device in federated learning, then comprehensive data utilization is achieved, but computation and communication overhead increases significantly
Solution Approach 1:
The patent divides the complete neural network into multiple segments distributed across different user devices. Each device performs computation only on its assigned segment using local data, significantly reducing computation overhead and energy consumption while still utilizing local data effectively.
Solution Approach 2:
The patent extracts and transmits only the necessary gradient updates from each device's local training to the ring topology, rather than transmitting full model parameters or raw data. This extraction approach minimizes communication overhead while preserving data utilization benefits.
4Reliability
If additional privacy protection mechanisms are added to federated learning, then privacy security is enhanced, but system efficiency and model performance deteriorate
Solution Approach 1:
The ring topology communication structure acts as an intermediary that aggregates gradient updates from multiple devices without requiring any single device to expose its raw data or full model. This intermediary structure provides inherent privacy protection while maintaining system efficiency.
Data Source
AI summary
A model training method based on adaptive split learning-federated learning includes: each user terminal uploading device information to the server and the server allocating a propagation step length and a aggregation weight to each user terminal; each user terminal obtaining a current-round global model from the server and taking itself as a start node of a ring topology to perform local joint processing for a preset number of times to obtain a locally-updated model parameter of the start node with respect to current-round training; each user terminal uploading the locally-updated model parameter for the current-round training to the server for aggregation and obtaining a current-round updated global model; and the server determining whether the current-round updated global model meets a convergence condition, if not, performing next-round training, or if yes, determining the current-round updated global model as a trained face recognition model.