基于张量分解与网络权重的终端数据同分布识别算法

Tensor Decomposition and Neural Network Weights Based Terminal Data Distribution Identification

联邦学习协调多个异地终端使用当地用户数据训练神经网络,并用服务器收集已训练的网络权重更新全局神经网络。联邦学习无须收集用户数据,保护了用户隐私,且利用了用户数据训练网络。然而,由于用户偏好的差异,用户数据常不满足同一分布,导致无法使用一个全局神经网络拟合所有数据。为了让服务器能依据终端训练的权重识别数据分布,用不同网络拟合不同分布的终端,提出一种基于张量分解的算法仅使用网络权重识别含相同数据分布的终端。将终端所训练的神经网络权重表示为高维度张量,并使用张量分解自适应地寻找潜在最优子空间,学习每个权重更具辨识性的新表征,对新表征聚类以识别出具有相同数据分布的终端。使用多个数据集验证了该算法对终端数据分布一致性识别的高效性。

Federated learning collaboratively trains neural networks with remote terminals and thereof user data,then collects trained weights to update global network in the server.Such learning strategy avoids user privacy leakage resulting by collecting data,and avails these data for network training.However,due to the difference amongst user preferences,terminals usually have non-identical data distributions,leading to the infeasibility of fitting all data with single global network.In order to discriminate terminals in terms of thereof data distributions with the trained weights,and to fit identical data with the same global network,this paper presents a tensor decomposition based method to identify the homogenous terminals with identical data distributions.The proposed method stacks the trained network weights into high-dimensional tensor,and transforms weights into potentially optimal subspace by tensor decomposition to obtain discriminative representations of weights.Further,the method analyzes the representations to identify terminals with identical distributions.