Remote physiological measurement gained wide attention, while it requires collecting users' privacy-sensitive information, and existing contactless measurements still rely on labeled client data. This presents challenges when we want to further update real-world deployed models with numerous user data lacking labels. To resolve these challenges, we instantiate a new protocol called Federated Unsupervised Domain Generalization (FUDG) in this work. Subsequently, the Federated Heterogeneous Unsupervised Generalization (FedHUG) framework is proposed and consists of: (1) Minimal Bias Aggregation module dynamically adjusts aggregation weights based on prior-driven bias evaluation to cope with heterogeneous non-IID features from multiple domains. (2) The Global Distribution-aware Learning Controller parameterizes the label distribution and dynamically manipulates client-specific training strategies, thereby mitigating the server-client label distribution skew and long-tail issue. The proposal shows superior performance across state-of-the-art techniques in estimation with either RGB video or mmWave radar. The code will be released.
Figures & tables
Figure 1: (a) Overview of FUDG. We take the video-based physiological estimation as an example. (b) Domain Skew. Significant feature shifts under varying user attributes and environmental factors.(c) Label Skew. Inherent long-tail and imbalanced distribution of phys-labels across datasets (i.e., PURE Song et al. [2021] , BUAA Xi et al. [2020] , VIPL Lu et al. [2021] , MMPD Tang et al. [2023] ).
Figure 2: RGB representation construction and physiologically consistent augmentation. (a) Production of an STMap from a facial RGB video using region-wise color traces. (b) Unified spatio-temporal augmentation, illustrated with an STMap: temporal sliding introduces a phase shift δt , while spatial permutation disrupts structural and environmental noise. The same temporal and spatial operations are applied to mmWave range matrices.
Figure 3: Illustration of three different aim domains. Here, the dashed circle’s center represents the semantic features, and the closer a domain is to the center, the less interference it contains.
Figure 4: Minimal bias evaluation via SSM generation. (a) The robust model yields highly consistent translation-invariant SSMs despite the phase shift between the original ( xi ) and augmented ( ai ) views. (b) The biased model succumbs to domain-specific noise, resulting in corrupted and inconsistent SSM textures. This visual discrepancy enables the score s to accurately quantify model bias.
Figure 5: Illustration of the distribution of s in different datasets.
VIPL
PURE
BUAA
MMPD
Method
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
CHROM De Haan and Jeanne [2013]
11.44
15.19
16.97
0.28
9.79
11.96
12.76
0.37
6.09
8.03
8.29
0.51
13.63
16.23
18.75
0.08
POS Wang et al. [2016]
14.59
18.11
21.26
0.19
9.82
12.85
13.44
0.34
5.04
6.54
7.12
0.63
13.84
17.07
19.18
0.15
PhysNet + Yu et al. [2019]
13.53
15.13
16.46
0.35
18.12
22.71
24.32
0.43
7.21
7.40
8.12
0.58
19.11
24.17
26.29
0.04
Physformer++ + Yu et al. [2023]
12.07
14.34
15.38
0.32
15.48
19.50
24.63
0.46
6.52
8.47
8.51
0.55
16.37
19.65
21.82
0.07
RhythmFormer + Zou et al. [2025]
10.22
12.04
12.98
0.39
6.47
8.02
9.45
0.61
4.78
6.03
6.16
0.64
13.24
15.98
16.53
0.10
Table 1: HR estimation results under FUDG protocol in the rPPG task. ‘Base’ and ‘ + ’ refer to the pre-trained model that is applied directly to the target without unsupervised training. ‘ ∗ ’ means it uses the same unsupervised loss and backbone as FedHUG. The bold texts show the significantly (p-value < 0.05) best result within each column with the paired-t test.
VIPL
PURE
BUAA
MMPD
Method
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
FedAvg McMahan et al. [2017]
6.86
8.98
9.39
0.52
5.61
7.71
8.09
0.57
3.80
4.65
4.89
0.82
9.79
11.24
11.79
0.13
FedProx ∗ Li et al. [2020]
6.78
8.73
9.26
0.54
5.51
7.66
7.94
0.59
3.35
4.81
4.81
0.84
9.69
11.07
11.60
0.14
FedMut ∗ Hu et al. [2024a]
6.82
8.52
9.41
0.48
5.88
8.25
8.96
0.58
3.68
4.74
5.05
0.83
9.77
11.28
11.93
0.11
FedU2 Liao et al. [2024]
6.89
8.48
9.20
0.53
5.78
7.41
8.15
0.59
3.49
4.61
4.72
0.82
9.85
10.88
11.59
0.14
FedGaLA Pourpanah et al. [2025]
6.61
8.38
9.13
0.55
5.61
7.29
7.74
0.56
3.29
4.60
4.77
0.85
9.70
10.69
11.55
0.16
Table 2: Results on source domains. The model is pre-trained on SCAMPS and federated unsupervised trained on four datasets.
Flat & Unobstructed
Flat & Congested
Bumpy & Congested
Method
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
MAE ↓
SD ↓
RMSE ↓
p ↑
Base
11.21
14.66
15.91
0.09
11.45
14.35
15.27
0.12
11.56
14.98
15.73
0.10
FedAvg
10.03
12.74
13.07
0.19
9.89
12.60
12.56
0.15
10.21
12.78
12.84
0.16
FedProx
9.79
12.60
12.85
0.23
9.54
12.42
13.66
0.21
9.95
12.83
12.90
0.19
FedMut
9.90
11.87
12.67
0.22
10.19
11.97
13.12
0.18
9.93
12.41
12.49
0.23
FedU2
9.87
11.49
11.54
0.25
9.97
10.66
10.93
0.24
10.22
11.06
12.41
0.23
Table 3: HR estimation results with mmWave under FUDG.
Figure 6: Illustration of the distribution of s in PURE before and after applying the MBA method (a), and MAE at different HR intervals before and after the use of GDLC (b). MAE on MMPD across communication rounds under the FUDG protocol (c) . The MAE variation across different communication rounds of the MMPD dataset under the FUDG protocol, the impact of the γ and σ on the model’s performance (d). BUAA’s feature distribution on unit sphere (e).
Unsupervised federated learning (UFL) enables privacy-preserving distributed training without data labeling, yet practical deployment remains challenging due to non-IID data, high computational and communication costs at edge devices, and sensitivity to communication noise. We propose FedUHD, the first UFL framework based on Hyperdimensional Computing (HDC). On the client side, FedUHD employs kNN-based cluster hypervector removal to mitigate non-IID effects by filtering detrimental local outliers. On the server side, cluster-aware HDC aggregation leverages cluster-level statistics to stabilize learning across heterogeneous clients. To further improve efficiency, we design a compute-in-memory (CIM) accelerator based on a novel phase-change memory (PCM) device, integrated with lightweight ASIC digital modules to execute the client-side HDC pipeline within the accelerator. The intrinsic robustness of HDC to low precision and device variations enables efficient mapping onto analog PCM crossbars, exploiting massive parallelism while minimizing data movement. Experimental results show that FedUHD achieves comparable accuracy to state-of-the-art neural network-based UFL methods across all datasets. On HAR and CIFAR10/100, FedUHD delivers an average 2,239x speedup and 1,542x higher energy efficiency on GPU. In addition, FedUHD reduces communication cost by up to 176x on HAR and CIFAR10/100 and demonstrates greater robustness than Orchestra under communication noise. Compared to GPU implementation of FedUHD, the proposed PCM-based accelerator provides an additional 4.07x speedup and three orders of magnitude higher energy efficiency on average. Furthermore, the results demonstrate the benefit of PCM over RRAM as a CIM substrate.
The federated learning (FL) paradigm fosters distributed pervasive computing combined with artificial intelligence techniques, allowing for optimized data usage and improved mitigation of privacy concerns. Indeed, model training occurs on the client's local devices, and model parameters are subsequently shared with a centralized server. However, there is a need to find a tradeoff between models' personalization and generalization capabilities. In this paper, we design and implement several testing scenarios devoted to evaluating and comparing the centralized, local, and federated paradigm performances. We also design and implement a scenario that emulates a change in clients' data. We then present experimental results of the FedAvg algorithm applied to the Human Activity Recognition (HAR) domain to understand the trade-off between personalized and generalized accuracy. Results show that, although FedAvg confirms a higher degree of personalization capabilities while keeping a high degree of generalization with respect to the traditional centralized learning, this result is not so obvious under stressful conditions, such as when varying class distribution over clients.
Andrea De Luna, Susanna Peretti, Chiara Contoli +1
Federated learning enables collaborative model training without sharing raw data, but its performance can degrade substantially under heterogeneous client data distributions. A single global model often cannot satisfy diverse client requirements, so personalized federated learning has therefore been explored to improve client specific performance while preserving global generalization. Existing PFL methods often face a fundamental tradeoff in which stronger global sharing can undermine local specialization, whereas stronger local adaptation can lead to overfitting under limited data, label imbalance, and missing class scenarios. In this work, we propose PGFedSplit, a personalized federated learning framework that improves both personalization and global generalization under severe client heterogeneity. PGFedSplit adopts a split architecture and performs adaptive aggregation scheduling tailored to the roles of different model components, enabling stable knowledge sharing while maintaining client specific adaptation. Each client further leverages a mixture of locally extracted representations and synthetic representations generated from server side Gaussian statistics, improving robustness under label imbalance and missing class conditions. Extensive experiments on Fashion MNIST, CIFAR 10, CIFAR 100, and Tiny ImageNet demonstrate consistent improvements over state of the art PFL methods, with stable convergence and superior personalization in highly heterogeneous settings.
Yunseok Kang, Jaeyoung Song
Department of Electronics Engineering, Pusan National University, Republic of Korea