Geometric Knowledge-Guided Localized Global Distribution Alignment for Federated Learning

Geometric Knowledge-Guided Localized Global Distribution Alignment for Federated Learning
Yanbiao Ma*, Wei Dai*, Wenke Huang†, Jiayi Chen
Accepted to CVPR 2025 🔗Link

📝 Abstract

Data heterogeneity in federated learning, characterized by a significant misalignment between local and global distributions, leads to divergent local optimization directions and hinders global model training.Existing studies mainly focus on optimizing local updates or global aggregation, but these indirect approaches demonstrate instability when handling highly heterogeneous data distributions, especially in scenarios where label skew and domain skew coexist.To address this, we propose a geometry-guided data generation method that centers on simulating the global embedding distribution locally. We first introduce the concept of the geometric shape of an embedding distribution and then address the challenge of obtaining global geometric shapes under privacy constraints. Subsequently, we propose GGEUR, which leverages global geometric shapes to guide the generation of new samples, enabling a closer approximation to the ideal global distribution.In single-domain scenarios, we augment samples based on global geometric shapes to enhance model generalization;in multi-domain scenarios, we further employ class prototypes to simulate the global distribution across domains.Extensive experimental results demonstrate that our method significantly enhances the performance of existing approaches in handling highly heterogeneous data, including scenarios with label skew, domain skew, and their coexistence.

📂 New Dataset: Office-Home-LDS

Dataset & Constructor: 👉 📥 Huggingface

The dataset is organized as follows:

Office-Home-LDS/
├── data/
│   └── Office-Home.zip        #  Original raw dataset (compressed)
├── new_dataset/               #  Processed datasets based on different settings
│   ├── Office-Home-0.1.zip    #  Split with Dirichlet α = 0.1 (compressed)
│   ├── Office-Home-0.5.zip    #  Split with Dirichlet α = 0.5 (compressed)
│   └── Office-Home-0.05.zip   #  Split with Dirichlet α = 0.05 (compressed)
├── Dataset-Office-Home-LDS.py #  Python script for processing and splitting original raw dataset
└── README.md                  #  Project documentation

🌐 Environment

Provide two options

(1): Download dependency

conda create -n GGEUR python=3.9
conda activate GGEUR
pip install -r requirements.txt

(2): Download environment (Windows)

conda env create -f environment.yml
conda activate GGEUR

🚀 Single Domain

🗂️ 1. Dataset Partitioning

CIFAR 10 & 100 dataset

Dataset index parsing:

The CIFAR dataset is sorted by subsets of the same class, so the indexing process is directly arranged by the number of classes.

Optional Parameters

Parameter	Type	Description
`num_clients`	integer	Number of clients
`alpha`	float	Dirichlet coefficient
`min_require_size`	integer	Minimum allocation size

Run Script

python data_distribution_CIFAR-10.py
python data_distribution_CIFAR-100.py
python data_batch2index_images.py  # Check the partitioning

Output

{dataset_name}/context/alpha={alpha}_class_counts.txt         #  Total count of each class after partitioning (for validation)
{dataset_name}/context/alpha={alpha}_class_indices.txt        #  Class indices based on partitioning
{dataset_name}/context/alpha={alpha}_client_indices.txt       #  Data indices assigned to each client
{dataset_name}/context/alpha={alpha}_client_class_distribution.txt #  Class distribution across clients
{dataset_name}/images/alpha={alpha}_label_distribution_heatmap.png #  Heatmap showing the number of each class assigned to each client

TinyImageNet dataset

Dataset Structure

📌 The TinyImageNet dataset is derived from 200 classes of ImageNet.
⚠️ The class order is not linear, so reconstruction is required.

TinyImageNet Structure:
│
├── train/
│   └── 🗂️ n01443537/   # Class folder
│       └──  images/  # Training images for the class
│           └──  image_001.JPEG
│           └──  image_002.JPEG
│           └── ...
├── val/
│   ├──  images/       # Validation images
│   └──  val_annotations.txt  # Mapping between images and class IDs
│
├── test/
│    └──  images/       # Test images
│
├── wnids.txt         # Class ID list
│
└── words.txt         # Class label list

👉 Since the test set has no labels, the validation set will be used as the test set.
👉 The training set follows the same folder structure as ImageNet.
👉 Validation images are stored in val/images, and the label mapping is in val/val_annotations.txt. 👉 The script Reorganized_TinyImageNet_Val.py aligns the validation set structure with the training set:

│
├── train/
│   └── 🗂️ n01443537/   # Class folder
│       └──  images/  # Training images for the class
│           └──  image_001.JPEG
│           └──  image_002.JPEG
│           └── ...
├── new_val/
│   └── 🗂️ n01443537/   # Class folder
│       └──  images/  # Validation images for the class
│           └──  image_001.JPEG
│           └──  image_002.JPEG
│           └── ...

💡 After this:
Label IDs (n + 8 digits) will be indexed.For consistency, the n prefix will be removed, and IDs will be sorted to define the class order.

📌 This is handled by:

📝 data_distribution_TinyImageNet.py (training set partitioning)
📝 TinyImageNet_Val.py (validation set partitioning) At this point, we have completed the reconstruction of the TinyImageNet dataset, laying the groundwork for the subsequent indexing and partitioning tasks.

Parameters

Parameter	Type	Description
`num_clients`	integer	Number of clients
`alpha`	float	Dirichlet coefficient
`min_require_size`	integer	Minimum allocation size

Run Scripts

python reorganized_TinyImageNet_val.py         # Reorganize validation set
python TinyImageNet_val_index.py               # Convert validation set indices
python data_distribution_TinyImageNet.py       #  Partition training set
python TinyImageNet_val_index_tag_img_matching_test.py  # Validate processed validation set
python TinyImageNet_val_features.py            # Extract validation set features

Output

{dataset_name}/context/alpha={alpha}_class_counts.txt         # Total count of each class after partitioning (for validation)
{dataset_name}/context/alpha={alpha}_class_indices.txt        # Class indices based on partitioning
{dataset_name}/context/alpha={alpha}_client_indices.txt       # Data indices assigned to each client
{dataset_name}/context/alpha={alpha}_client_class_distribution.txt # Class distribution across clients
{dataset_name}/images/alpha={alpha}_label_distribution_heatmap.png # Heatmap showing the number of each class assigned to each client
{dataset_name}/context/class_map.txt                          # Mapping between class index and class ID (training set)
{dataset_name}/val_context/class_map.txt                      # Mapping between class index and class ID (validation set)
{dataset_name}/val_context/val_indices.npy                    # Validation set indices
{dataset_name}/val_context/val_labels.npy                     # Labels corresponding to validation set indices
{dataset_name}/class_{class_label}_val_indices.npy            # Indices of each class in validation set

✅ 2. Cross Indexing

We have obtained the data indexes for the CIFAR-10, CIFAR-100, and TinyImageNet datasets, including the data indexes for each client and each class.

By performing cross-indexing between the two, we can generate the data indexes for each class under each client.

Taking CIFAR-10 as an example: 10 clients × 10 classes = 100 cross-index files

Parameters

Parameter	Type	Description
`dataset`	string	Dataset name (CIFAR-10, CIFAR-100, TinyImageNet)
`alpha`	float	Dirichlet coefficient

Run Script

python client_class_cross_index.py

Output

{dataset_name}/client_class_indices/alpha={alpha}_{dataset}_client_{client_id}_class_{class_id}_indices.npy

✅ 3. Feature Extraction

We have obtained the index files of each class for each client in the three datasets.
Using CLIP as the backbone, we extract features for each index file and generate the corresponding feature and label files.

Parameters

Parameter	Type	Description
`dataset`	string	Dataset name (CIFAR-10, CIFAR-100, TinyImageNet)
`alpha`	float	Dirichlet coefficient

Run Script

python client_class_clip_features2tensor.py

Output

{dataset_name}/features/initial/alpha={alpha}_class_{class_idx}_client_{client_idx}/final_embeddings.npy
{dataset_name}/features/initial/alpha={alpha}_class_{class_idx}_client_{client_idx}/labels.npy

✅ 4. Global Distribution

🔸 4.1 Global Distribution of Client Composition

The file alpha={alpha}_client_class_distribution.txt generated in Section 1 contains the class distribution for each client.
Clients are sorted based on the number of classes, and a set of clients approximating the global distribution is determined according to a threshold ratio.

Parameters

Parameter	Type	Description
`threshold`	float	Threshold
`dataset`	string	Dataset name (CIFAR-10, CIFAR-100, TinyImageNet)
`alpha`	float	Dirichlet coefficient

Run Script

python client-guided_set.py

Output

{dataset_name}/context/alpha={alpha}_selected_clients_for_each_class.txt

🔸 4.2 Client Set Features (Optional)

In Section 3, we obtained feature files for each class for each client.
In Section 4.1, we obtained the client sets that approximate the global distribution for each class.
We now extract the corresponding feature files from the former using the latter.

Parameters

Parameter	Type	Description
`dataset`	string	Dataset name (CIFAR-10, CIFAR-100, TinyImageNet)
`alpha`	float	Dirichlet coefficient

Run Script

python client-guided_clip_tensor.py

🔸 4.3 Representation of Global Distribution

If the client set contains only one client, the covariance matrix is calculated directly from the feature matrix.
If the client set contains multiple clients, the covariance matrix is calculated separately for each client’s feature matrix, and the resulting covariance matrices are aggregated to form the final aggregated covariance matrix.

Parameters

Parameter	Type	Description
`dataset`	string	Dataset name (CIFAR-10, CIFAR-100, TinyImageNet)
`alpha`	float	Dirichlet coefficient

Run Script

python clip_tensor2aggregate_covariance_matrix.py

Output

{dataset_name}/features/alpha={alpha}_cov_matrix/class_{idx}_cov_matrix.npy

✅ 5. Geometry-Guided Data Augmentation

Now that the global distribution framework (covariance matrix) has been obtained, the covariance matrix is decomposed to derive eigenvalues and eigenvectors (geometric directions).
These geometric directions are used to guide the augmentation of raw samples from the client set.

Parameters

Parameter	Type	Description
`dataset`	string	Dataset name (CIFAR-10, CIFAR-100, TinyImageNet)
`alpha`	float	Dirichlet coefficient

Run Script

python cov_matrix_generate_features.py

Output

{dataset_name}/features/alpha={alpha}_complete/final_embeddings_filled.npy
{dataset_name}/features/alpha={alpha}_complete/labels_filled.npy

✅ 6. Single Client Training

Under a non-federated learning architecture, train both the original and augmented samples locally and compare the performance differences.

Parameters

Parameter	Type	Description
`alpha`	float	Dirichlet coefficient
`client_idx`	integer	Client ID
`batch_size`	integer	Batch size
`learning_rate`	float	Learning rate
`num_epochs`	integer	Number of training epochs

Run Script

python MLP_10.py   # Train on CIFAR-10 dataset
python MLP_100.py  # Train on CIFAR-100 dataset
python MLP_200.py  # Train on TinyImageNet dataset

✅ 7. Federated Architecture FedAvg Training

Under a basic federated learning architecture (FedAvg with simple averaging), train models using both the original and augmented samples and compare the performance differences.

Parameters

Parameter	Type	Description
`alpha`	float	Dirichlet coefficient
`communication_rounds`	integer	Number of communication rounds
`local_epochs`	integer	Number of local training epochs
`client_count`	integer	Number of clients
`batch_size`	integer	Batch size
`learning_rate`	float	Learning rate

Run Script

python FL_MLP_10.py   # Train on CIFAR-10 dataset
python FL_MLP_100.py  # Train on CIFAR-100 dataset
python FL_MLP_200.py  # Train on TinyImageNet dataset

🚀 Cross-Domain Scenarios

We conducted experiments on the Digits, PACS, and Office-Caltech-10 datasets, and proposed a new dataset called Office-Home-LDS.
Following standard cross-domain work:

Each domain is assigned one client.
Data is randomly partitioned based on ratio used in prior work.
For datasets without explicit training and test splits, we follow the ratio used in prior work.

🏷️ Digits

📌1. Dataset Overview

The Digits dataset contains data from four different domains, representing digits from 0 to 9:

MNIST – 🖋️ Handwritten digits from different individuals (0-9)
SVHN – 🏠 Street View House Numbers (0-9)
USPS – 📬 Postal Service digits (0-9)
SynthDigits – 🎨 Synthetic RGB digits (0-9)

📂 2. Dataset Partitioning

Following the standard settings for federated cross-domain work:

Each domain is assigned one client.
Data is randomly partitioned based on ratio used in prior work.
For datasets without explicit training and test splits, we follow the ratio used in prior work.

Run Script

python data_distribution_digits.py

Output

output_indices/{domain_name}/client_{client_id}_indices.npy         # Indices assigned to the client within the domain
output_indices/{domain_name}/combined_class_indices.npy             # Combined class indices within the domain
output_indices/client_combined_class_distribution.txt              # Class distribution per client within each domain
output_indices/dataset_report.txt                                  # Total samples assigned to each client (for validation)

✅ 3. Cross Indexing

We have obtained client indices and class indices for each domain.
By performing cross-indexing, we can generate class-specific indices for each client.

Run Script

python client_class_cross_index.py

Output

output_client_class_indices/{domain_name}/client_{client_id}_class_{0~9}_indices.npy

✅ 4. Training Set Feature Extraction

Using CLIP as the backbone, we extract features and labels for each client-class index file.

Parameters

Parameter	Type	Description
`datasets`	string	Domain name (MNIST, SVHN, USPS, SynthDigits)
`report_file`	string	Number of communication rounds

Run Script

python train_client_class_clip_features2tensor.py

Output

clip_features/{domain_name}/client_{client_id}_class_{0~9}_original_features.npy
clip_features/{domain_name}/client_{client_id}_class_{0~9}_labels.npy

✅ 5. Test Set Feature Extraction

We extract features and labels for the test set using CLIP as the backbone.

Parameters

Parameter	Type	Description
`datasets`	string	Domain name (MNIST, SVHN, USPS, SynthDigits)

Run Script

python test_clip_features2tensor.py

Output

clip_test_features/{domain_name}/{domain_name}_test_features.npy
clip_test_features/{domain_name}/{domain_name}_test_labels.npy

✅ 6. Prototype Extraction

Using the client-class index files, we extract prototypes for each class.

Run Script

python prototype_clip_features2tensor.py

Output

clip_prototypes/{domain_name}/client_{client_id}_class_{0~9}_prototype.npy

✅ 7. Geometric Direction

From the perspective of the manifold space, cross-domain differences are caused by shifts in class distribution, but the geometric structure remains unchanged.
Thus, we can use the combined features from multiple domains to represent the geometric structure.

Parameters

Parameter	Type	Description
`report_file`	string	File mapping domains to clients

Run Script

python clip_tensor2aggregate_covariance_matrix.py

Output

cov_matrix_output/class_{0~9}_cov_matrix.npy

✅ 8. Geometry-Guided Data Augmentation

We now have the geometric structure and class prototypes for multiple domains.
For data augmentation:

For samples within the same domain, apply domain-specific augmentation.
For samples outside the domain, augment based on class prototypes and geometric structure.

This allows the client to learn cross-domain features even from its own perspective, thereby effectively mitigating the bias caused by domain shift.

Run Script

python prototype_cov_matrix_generate_features.py

Output

argumented_clip_features/{domain_name}/client_{client_id}_class_{0~9}/final_embeddings_filled.npy
argumented_clip_features/{domain_name}/client_{client_id}_class_{0~9}/labels_filled.npy

✅ 9. Training Under Different Federated Architectures

We train both the original and augmented models under different federated architectures to compare performance:

Run Script

python FedAvg.py
python FedAvg_GGEUR.py
python FedNTD.py
python FedNTD_GGEUR.py
python FedOpt.py
python FedOpt_GGEUR.py
python FedProx.py
python FedProx_GGEUR.py
python MOON.py
python MOON_GGEUR.py
python FedDyn.py
python FedDyn_GGEUR.py
python FedProto.py
python FedProto_GGEUR.py
python SCAFFOLD.py
python SCAFFOLD_GGEUR.py

🌍 PACS and Office-Caltech-10

PACS and Office-Caltech-10 are small-sample datasets with fewer classification tasks:

✅ Fewer classification tasks – Low classification difficulty.
✅ High inter-class variation – Easier to distinguish between classes.
✅ Similar training and test sets – Improves accuracy implicitly.
✅ Balanced intra-domain classes – Following a proportional random split.

With the development of self-supervised learning, the challenges posed by cross-domain datasets like PACS and Office_Caltech_10 have been significantly weakened when using backbones such as CLIP and DINO.

Therefore, there is an urgent need for a more realistic and challenging dataset that better reflects real-world scenarios. To address this, we propose Office-Home-LDS — a dataset built upon the Office-Home dataset (65 classes) that incorporates both cross-domain(domain skew) and data heterogeneity(label skew). This clearly aligns better with real-world conditions.

🏷️ PACS

📌 1. Dataset Overview

The PACS dataset contains data from four different domains, with seven categories:
Dog, Elephant, Giraffe, Guitar, Horse, House, Person

P (Photo) – 📷 Real-world photos
A (Art Painting) – 🎨 Artistic paintings
C (Cartoon) – 🐾 Cartoon images
S (Sketch) – ✍️ Sketches

📂 2. Dataset Partitioning

Following standard cross-domain work:

Each domain is assigned one client.
Data is randomly partitioned based on ratio used in prior work.
For datasets without explicit training and test splits, we follow the ratio used in prior work.

Run Script

python data_distribution_digits.py

Output

./output_indices/{domain_name}/train_train_indices.npy          # Training set indices for the domain
./output_indices/{domain_name}/test_test_indices.npy            # Test set indices for the domain
./output_indices/{domain_name}/client_{client_id}_indices.npy   # Client-assigned indices for the domain
./output_indices/{domain_name}/class_indices.npy                # Combined class indices within the domain
./output_indices/client_combined_class_distribution.txt         # Class distribution per client within each domain

✅ 3. Cross Indexing

We have obtained the client indices and class indices for each domain.
By performing cross-indexing, we can generate class-specific indices for each client.

Run Script

python client_class_cross_index.py

Output

output_client_class_indices/{domain_name}/client_{client_id}_class_{0~6}_indices.npy

✅ 4. Training Set Feature Extraction

We have obtained class-specific index files for each client in the four domains.
Using CLIP as the backbone, we extract features for each index file and generate the corresponding feature and label files.

Run Script

python train_client_class_clip_features2tensor.py

Output

clip_pacs_train_features/{domain_name}/client_{client_id}_class_{0~6}_original_features.npy
clip_pacs_train_features/{domain_name}/client_{client_id}_class_{0~6}_labels.npy

✅ 5. Test Set Feature Extraction

We extract features and labels for the test set using CLIP as the backbone.

Run Script

python test_clip_features2tensor.py

Output

clip_test_features/{domain_name}/{domain_name}_test_features.npy
clip_test_features/{domain_name}/{domain_name}_test_labels.npy

✅ 6. Training Under Different Federated Architectures

We train both the original and augmented models under different federated architectures to compare performance:

Run Script

python FedAvg.py
python FedNTD.py
python FedOpt.py
python FedProx.py
python MOON.py
python FedDyn.py
python FedProto.py
python SCAFFOLD.py

🏷️ Office-Caltech-10

📌 1. Dataset Overview

The Office-Caltech-10 dataset contains data from four different domains, representing 10 categories:
Headphones, Keyboard, Laptop, Monitor, Mouse, Mug, Projector, Bike

A (Amazon) – 🛒 Product images from Amazon
W (Webcam) – 📷 Images captured from a webcam
D (DSLR) – 📸 Images captured from a DSLR camera
C (Caltech-256) – 🎯 Selected images from the Caltech-256 dataset

📂 2. Dataset Partitioning

Following standard federated cross-domain work:

Each domain is assigned one client.
Data is randomly partitioned based on a predefined ratio.
For datasets without explicit training and test splits, we follow the ratio used in prior work.

Run Script

python data_distribution_office_caltech_10.py

Output

output_indices/{domain_name}/train_train_indices.npy         # Training set indices for the domain
output_indices/{domain_name}/test_test_indices.npy           # Test set indices for the domain
output_indices/{domain_name}/client_{client_id}_indices.npy  # Client-assigned indices for the domain
output_indices/{domain_name}/class_indices.npy               # Combined class indices within the domain
output_indices/client_combined_class_distribution.txt        # Class distribution per client within each domain

✅ 3. Cross Indexing

We have obtained client indices and class indices for each domain.
By performing cross-indexing, we can generate class-specific indices for each client.

Run Script

python client_class_cross_index.py

Output

output_client_class_indices/{domain_name}/client_{client_id}_class_{0~64}_indices.npy

✅ 4. Training Set Feature Extraction

We have obtained class-specific index files for each client in the four domains.
Using CLIP as the backbone, we extract features for each index file and generate the corresponding feature and label files.

Run Script

python train_client_class_clip_features2tensor.py

Output

clip_office_home_train_features/{domain_name}/client_{client_id}_class_{0~64}_original_features.npy
clip_office_home_train_features/{domain_name}/client_{client_id}_class_{0~64}_labels.npy

✅ 5. Test Set Feature Extraction

We extract features and labels for the test set using CLIP as the backbone.

Run Script

python test_clip_features2tensor.py

Output

clip_test_features/{domain_name}/{domain_name}_test_features.npy
clip_test_features/{domain_name}/{domain_name}_test_labels.npy

✅ 6. Training Under Different Federated Architectures

We train both the original and augmented models under different federated architectures to compare performance:

Run Script

python FedAvg.py
python FedNTD.py
python FedOpt.py
python FedProx.py
python MOON.py
python FedDyn.py
python FedProto.py
python SCAFFOLD联.py

🏷️ Office-Home-LDS

📌 1. Dataset Overview

The Office-Home-LDS dataset contains data from four different domains, covering 65 categories:

Art – 🎨 Hand-drawn, sketch, or oil painting-style images
Clipart – 🖼️ Cartoon and clipart images from the web
Product – 🛒 Product or merchandise display images
Real World – 🌍 Photographs captured from real-world scenarios

📂 2. Dataset Partitioning

Following standard federated cross-domain work:

Each client is assigned one domain.
Data is randomly partitioned based on the Dirichlet distribution within the training set.
For datasets without explicit training and test splits, we follow the ratio used in prior work.

Run Script

python data_distribution_Office_Home_LDS.py

Output

./output_indices/{domain_name}/train_train_indices.npy         # Training set indices for the domain
./output_indices/{domain_name}/test_test_indices.npy           # Test set indices for the domain
./output_indices/{domain_name}/client_{client_id}_indices.npy  # Client-assigned indices for the domain
./output_indices/{domain_name}/class_indices.npy               # Combined class indices within the domain
./output_indices/client_combined_class_distribution.txt        # Class distribution per client within each domain

✅ 3. Cross Indexing

We have obtained client indices and class indices for each domain.
By performing cross-indexing, we can generate class-specific indices for each client.

Run Script

python client_class_cross_index.py

Output

output_client_class_indices/{domain_name}/client_{client_id}_class_{0~64}_indices.npy

✅ 4. Training Set Feature Extraction

We have obtained class-specific index files for each client in the four domains.
Using CLIP as the backbone, we extract features for each index file and generate the corresponding feature and label files.

Run Script

python train_client_class_clip_features2tensor.py

Output

clip_office_home_train_features/{domain_name}/client_{client_id}_class_{0~64}_original_features.npy
clip_office_home_train_features/{domain_name}/client_{client_id}_class_{0~64}_labels.npy

✅ 5. Test Set Feature Extraction

We extract features and labels for the test set using CLIP as the backbone.

Run Script

python test_clip_features2tensor.py

Output

clip_test_features/{domain_name}/{domain_name}_test_features.npy
clip_test_features/{domain_name}/{domain_name}_test_labels.npy

✅ 6. Prototype Extraction

Using the client-class index files obtained earlier, we extract class prototypes for each client.

Run Script

python prototype_clip_features2tensor.py

Output

./office_home_prototypes/{domain_name}/client_{client_id}_class_{0~64}_prototype.npy

✅ 7. Geometric Direction

From the perspective of the manifold space, cross-domain differences are caused by shifts in class distribution, but the geometric structure remains unchanged.
Thus, we can use the combined features from multiple domains to represent the geometric structure.

Parameters

Parameter	Type	Description
`report_file`	string	File mapping domains to clients

Run Script

python clip_tensor2aggregate_covariance_matrix.py

Output

cov_matrix_output/class_{0~64}_cov_matrix.npy

✅ 8. Geometry-Guided Data Augmentation

We now have the geometric direction and class prototypes for multiple domains.
For data augmentation:

For samples within the same domain, apply a single-domain augmentation strategy.
For samples outside the domain, augment based on class prototypes and geometric directions.
This allows the client to learn cross-domain features and mitigate domain shift effectively.

Run Script

python prototype_cov_matrix_generate_features.py

Output

argumented_clip_features/{domain_name}/client_{client_id}_class_{0~64}/final_embeddings_filled.npy
argumented_clip_features/{domain_name}/client_{client_id}_class_{0~64}/labels_filled.npy

✅ 9. Training Under Different Federated Architectures

We train both the original and augmented models under different federated architectures to compare performance:

Run Script

python FedAvg.py
python FedAvg_GGEUR.py
python FedNTD.py
python FedNTD_GGEUR.py
python FedOpt.py
python FedOpt_GGEUR.py
python FedProx.py
python FedProx_GGEUR.py
python MOON.py
python MOON_GGEUR.py
python FedDyn.py
python FedDyn_GGEUR.py
python FedProto.py
python FedProto_GGEUR.py
python SCAFFOLD.py
python SCAFFOLD_GGEUR.py

Learning Trajectory (Updating...)

When I completed this project, I was a third-year undergraduate student. 🌿 I will share my learning trajectory and how to efficiently and comprehensively develop expertise in a specific field. 🌊 I believe that the most effective approach is to start by identifying high-quality review articles from top-tier journals. 📚 After forming a comprehensive understanding of the field, I recommend selecting detailed papers from the references cited in these outstanding reviews, focusing on those that align with the direction of our current work for in-depth study. 🔍 This process resembles a leaf with its veins hollowed out — our process of understanding is akin to a flood flowing through the leaf, with the central vein serving as the core from which knowledge selectively branches out in all directions. 🚀

2023Tpami "Deep Long-Tailed Learning: A Survey" Paper——Review on Long-Tailed Learning
2024Tpami "Vision-Language Models for Vision Tasks: A Survey" Paper & Github——Review on Vision-Language Large Models
2024Tpami "Federated Learning for Generalization, Robustness, Fairness: A Survey and Benchmark" Paper & Github——Review on Federated Learning
2021CVPR "Model-Contrastive Federated Learning" Paper & Github——MOON(Alignment of Local and Global Model Representations)
2022AAAI "FedProto: Federated Prototype Learning across Heterogeneous Clients" Paper——FedProto(Alignment of Local and Global Prototype Representations)
2023FGCS "FedProc: Prototypical contrastive federated learning on non-IID data" Paper——FedProc(Alignment of Local and Global Prototype Representations)
2020ICML "SCAFFOLD:Stochastic Controlled Averaging for Federated Learning" Paper——SCAFFOLD(Alignment of Local and Global Optimization Directions)
2021ICLR "FEDERATED LEARNING BASED ON DYNAMIC REGULARIZATION" Paper——FedDyn(Alignment of Local and Global Losses)
2022NeurIPS "Preservation of the Global Knowledge by Not-True Distillation in Federated Learning" Paper——FedNTD(Alignment of Unseen Local Losses with Global Losses)
2021ICLR "ADAPTIVE FEDERATED OPTIMIZATION" Paper——FedOpt(Server-Side Aggregation Optimization)
2024CVPR "Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity"Paper & Github——FedHEAL(Alignment of Local and Global Model Representations)
2023WACV "Federated Domain Generalization for Image Recognition via Cross-Client Style Transfer"Paper & Github——CCST(Alignment of Local and Global Optimization Directions)
2023TMC "FedFA: Federated Learning with Feature Anchors to Align Features and Classifiers for Heterogeneous Data" Paper——FedFA(Alignment of Features and Classifiers)
2024AAAI "CLIP-Guided Federated Learning on Heterogeneous and Long-Tailed Data" Paper——CLIP As Backbond For FL
2023CVPR "Rethinking Federated Learning with Domain Shift: A Prototype View" Paper & Github——Cross-Domain Prototype Loss Alignment
2023ICLR "FEDFA: FEDERATED FEATURE AUGMENTATION" Paper & Github——Class Prototype Gaussian Enhancement
2021ICLR "FEDMIX: APPROXIMATION OF MIXUP UNDER MEAN AUGMENTED FEDERATED LEARNING" Paper——Mixup For FL
2021PMLR "Data-Free Knowledge Distillation for Heterogeneous Federated Learning" Paper——Data-Free Knowledge Distillation For FL
2017ICML "Communication-Efficient Learning of Deep Networks from Decentralized Data" Paper——FedAvg(Average aggregation)
2025ICLR "Pursuing Better Decision Boundaries for Long-Tailed Object Detection via Category Information Amount" Paper——IGAM Loss(Revise decision boundaries)
2025Entropy "Trade-Offs Between Richness and Bias of Augmented Data in Long-Tailed Recognition" Paper——EIG(Effectiveness of distributed gain)

💡 Citation

If you find our work useful, please cite it using the following BibTeX format:

@article{ma2025geometric,
  title={Geometric Knowledge-Guided Localized Global Distribution Alignment for Federated Learning},
  author={Ma, Yanbiao and Dai, Wei and Huang, Wenke and Chen, Jiayi},
  journal={arXiv preprint arXiv:2503.06457},
  year={2025},
  url={https://arxiv.org/pdf/2503.06457}
}

📧 Contact

For any questions or help, feel welcome to write an email to
22012100039@stu.xidian.edu.cn or wdai@stu.xidian.edu.cn

Name		Name	Last commit message	Last commit date
Latest commit History 34 Commits
.idea		.idea
Multi Domain		Multi Domain
Single Domian		Single Domian
images		images
README.md		README.md
environment.yml		environment.yml
requirements.txt		requirements.txt

WeiDai-David/2025CVPR_GGEUR

Folders and files

Latest commit

History

Repository files navigation

Geometric Knowledge-Guided Localized Global Distribution Alignment for Federated Learning

📝 Abstract

📂 New Dataset: Office-Home-LDS

🌐 Environment

Provide two options

🚀 Single Domain

🗂️ 1. Dataset Partitioning

CIFAR 10 & 100 dataset

Dataset index parsing:

Optional Parameters

Run Script

Output

TinyImageNet dataset

Dataset Structure

Parameters

Run Scripts

Output

✅ 2. Cross Indexing

Parameters

Run Script

Output

✅ 3. Feature Extraction

Parameters

Run Script

Output

✅ 4. Global Distribution

🔸 4.1 Global Distribution of Client Composition

Parameters

Run Script

Output

🔸 4.2 Client Set Features (Optional)

Parameters

Run Script

🔸 4.3 Representation of Global Distribution

Parameters

Run Script

Output

✅ 5. Geometry-Guided Data Augmentation

Parameters

Run Script

Output

✅ 6. Single Client Training

Parameters

Run Script

✅ 7. Federated Architecture FedAvg Training

Parameters

Run Script

🚀 Cross-Domain Scenarios

🏷️ Digits

📌1. Dataset Overview

📂 2. Dataset Partitioning

Run Script

Output

✅ 3. Cross Indexing

Run Script

Output

✅ 4. Training Set Feature Extraction

Parameters

Run Script

Output

✅ 5. Test Set Feature Extraction

Parameters

Run Script

Output

✅ 6. Prototype Extraction

Run Script

Output

✅ 7. Geometric Direction

Parameters

Run Script

Output

✅ 8. Geometry-Guided Data Augmentation

Run Script

Output

✅ 9. Training Under Different Federated Architectures