Source-linked AI summary
HHHFL: Hierarchical Heterogeneous Horizontal Federated Learning for Electroencephalography
Dashan Gao, Ce Ju, Xiguang Wei, Yang Liu, Tianjian Chen, Qiang Yang
TL;DR
EEG classification over heterogeneous device data presents a challenge alongside privacy constraints. HHHFL maps device-specific data into a common embedding space and trains models through federated learning. The paper reports significant improvement over local models without federated learning.
Problem
Existing federated learning mainly focuses on homogeneous feature spaces, while EEG devices differ in channels and sampling rates and privacy concerns limit central data collection.
Source: §1 · pp. 1–2Method
HHHFL learns device-specific projections into a common embedding space and uses federated learning to train feature mappings and an EEG classifier without directly accessing raw EEG data.
Source: §1 · p. 2, §4.1 · p. 3, §4 · p. 3Results
HHHFL significantly outperforms models built locally without federated learning; on two- or three-device datasets it outperforms baselines in accuracy and convergence speed.
Source: §1 · p. 2, §5.3 · p. 4Takeaways & Limitations
The approach supports joint EEG classification across diverse devices while keeping raw EEG data on clients during federated training.
Source: §1 · p. 2Further privacy protection against training-time information leakage and membership-inference attacks is left to future work using secure multiparty computation and differential privacy.
Source: §6 · p. 5
Abstract
from arXiv · showhide
Electroencephalography (EEG) classification techniques have been widely studied for human behavior and emotion recognition tasks. But it is still a challenging issue since the data may vary from subject to subject, may change over time for the same subject, and maybe heterogeneous. Recent years, increasing privacy-preserving demands poses new challenges to this task. The data heterogeneity, as well as the privacy constraint of the EEG data, is not concerned in previous studies. To fill this gap, in this paper, we propose a heterogeneous federated learning approach to train machine learning models over heterogeneous EEG data, while preserving the data privacy of each party. To verify the effectiveness of our approach, we conduct experiments on a real-world EEG dataset, consisting of heterogeneous data collected from diverse devices. Our approach achieves consistent performance improvement on every task.
1 Introduction
EEG classification must contend with privacy constraints and device-specific differences in EEG features. HHHFL combines federated learning with shared feature mapping to address heterogeneous EEG data, and the paper reports significant improvement over local models without federated learning.
1 Introduction
- EEG-based models may require large amounts of data, but central collection raises privacy concerns and is constrained by data-protection requirements.The paper argues that privacy-preserving joint EEG analysis is significant when abundant data are difficult to collect for a single device.Source: §1 · p. 1
- Existing federated learning approaches mainly address homogeneous datasets in which parties share the same feature space.The introduction identifies heterogeneous EEG data and privacy constraints as gaps in prior work.Source: §1 · p. 2
- EEG data from different devices can vary in electrode number, position, and sampling rate, while each of numerous clients may have only limited labeled data.The paper says existing heterogeneous domain-adaptation approaches assume abundant labeled instances in source domains and would face significant accuracy drop in this setting.Source: §1 · p. 2
- HHHFL significantly outperforms models built locally without federated learning on the evaluated EEG tasks.The approach is designed for heterogeneous EEG signals collected from diverse devices.Source: §1 · p. 2
- HHHFL lets clients of the same device type collaboratively learn feature mappings into a common embedding space, then classifies the resulting homogeneous features.Each party acts as both source and target domain, and knowledge is accumulated and transferred among clients.Source: §1 · p. 2
2 Related Work
Prior EEG classification work includes several machine-learning approaches, while related federated and heterogeneous-domain methods address parts of the setting. The paper identifies a gap in combining heterogeneous EEG data from different devices with privacy-preserving federated classification.
2 Related Work
- Prior EEG classifiers include CNNs, MLPs and LSTMs with adaptive boosting, and heterogeneous classifiers using different base classifiers and feature extraction approaches [10] [11].CNNs were used to distinguish pathological from normal recordings [10].Source: §2 · p. 2
- The paper states that prior work had not attempted EEG classification over heterogeneous data collected from different devices while also considering privacy.It distinguishes this gap from prior study of different-device EEG data for the same task.Source: §2 · p. 2
- Federated multi-task learning treats each task separately [16], while a federated transfer-learning approach uses instance co-occurrence to build multiple correlated models.Symmetric transformations project source and target feature spaces into a domain-invariant subspace, with most approaches targeting one domain [18].Source: §2 · p. 2
3 Problem Statement
The problem is to classify EEG signals from private, high-dimensional sensor readings distributed across device-specific embedding submanifolds. The stated task is to detect whether the subject received a digit stimulus and thought about it.
3 Problem Statement
- Sensor readings from multiple brain-area sensors are high-dimensional and distributed on an embedding submanifold M in R^d.The paper describes d as large.Source: §3 · p. 3
- The readings are grouped into m sensor-reading clusters on submanifolds M_i in R^d_i, and the data are private and accessible only to users.The stated objective is a privacy-preserving classifier for the digit-stimulus task.Source: §3 · p. 3
- The classifier is intended to detect whether a subject received a digit-viewing stimulus and thought about it, using the corresponding EEG signal.The problem statement frames the signals as coming from device-specific clusters.Source: §3 · p. 3
4 Methodology
The methodology aligns device-specific EEG data in a common embedding space, optimizes classification and domain losses, and trains models through federated aggregation without direct access to raw training EEG data.
4 Methodology
- The approach uses federated learning to train a global model without direct access to raw training EEG data.The methodology identifies dimensionality and privacy as issues addressed in Sections 4.1 and 4.3, respectively.Source: §4 · p. 3, §4.3 · p. 4
4.1 Manifold Projection
- Each sensor cluster has its own projection P_i, approximated with a neural network, mapping its embedding submanifold into common embedding space N.The method requires raw EEG data to fall closely onto N.Source: §4.1 · p. 3
4.2 Loss
- The loss combines classification loss over EEG data and labels with a domain loss measuring distances between projected EEG distributions on N.The classification loss is cross-entropy, and the domain loss uses maximum mean discrepancy (MMD).Source: §4.2 · pp. 3–4
4.3 Federated learning
- In each round, the server aggregates feature-mapping models across clients of the same device and EEG classifiers across all clients, then returns updated global models.Clients update those models using local data on their projection manifolds until convergence.Source: §4.3 · p. 4
5 Experiments
Experiments on heterogeneous MindBigData device data compare HHHFL with device-specific baselines. HHHFL outperforms the baselines in accuracy and convergence speed across two- or three-device datasets.
5.1 MindBigData
- MindBigData experiments use EEG signals from MU, MW, and EP devices with different channel counts and sampling frequencies for stimulus classification.MU has 4 channels at 220 Hz, MW has 1 channel at 512 Hz, and EP has 14 channels at 128 Hz.Source: §5.1 · p. 4, §5 · p. 4
5.2 Evaluation Metric and Methods
- Accuracy is the evaluation metric; HHHFL is tested on each pair and on all three device datasets against a classifier trained separately on each device.The separate device classifiers provide three baselines.Source: §5.2 · p. 4
5.3 Implementation Details and Result Analysis
- Each device uses a CNN and fully connected private network, with a 10-dimensional common embedding; HFL is implemented with PySyft.Input dimensions are 512 for MU, 440 for EP, and 1024 for MW.Source: §5.3 · p. 4
- HHHFL on two- or three-device datasets outperforms the baselines in accuracy and convergence speed.Adding a third device yields a slight accuracy improvement on MW and EP.Source: §5.3 · p. 4, Figure 3 · p. 5, Figure 4 · p. 5
6 Future Work
The paper plans to augment privacy preservation with secure multiparty computation and differential privacy.
6 Future Work
- Future work plans to add secure multiparty computation and differential privacy to augment privacy preservation.The proposed additions address information leakage during model training and membership-inference attacks from adversarial model queriers.Source: §6 · p. 5
- Secure multiparty computation is proposed to prevent information leakage during model training.Source: §6 · p. 5
- Differential privacy is planned to prevent membership-inference attacks by adversarial model queriers.Source: §6 · p. 5
A Hierarchical Architecture
Figure 6 depicts HHHFL as a hierarchy built around device data blocks used in typical horizontal federated learning.
A Hierarchical Architecture
- The figure distinguishes two device-data blocks in typical horizontal federated learning from a yellow block placed hierarchically above them.Source: Figure 6 · p. 7
- Red and green blocks represent the two kinds of device data used in typical horizontal federated learning.Source: Figure 6 · p. 7
- The yellow block's hierarchical relation to the red and green blocks gives HHHFL its hierarchical architecture.Source: Figure 6 · p. 7