DisAgg: Distributed Aggregators for Efficient Secure Aggregation in Federated Learning

Hardware Requirements

CPU with 10+ cores, Nvidia GPU with at least 10 GB VRAM and minimum CUDA 12.2 support, at least 128 GB RAM, and 0.5 TB disk space.

Reference hardware: All experiments were conducted on a single machine using an Intel(R) Xeon(R) Gold 5220 @ 2.20GHz with 128GB RAM, and an NVIDIA GeForce RTX 2080 Ti with 12GB VRAM.

Installation

A virtual environment with Python 3.10 installed and bash as the default terminal is recommended.

Note: NumPy <2.x is required (breaking changes in NumPy 2.x).

# install as an editable package to allow modifying constants.py
pip install -e .

Confirm installation by executing: pip show disagg

Datasets

• MNIST / CIFAR-10: Downloaded automatically via torchvision on first run. No manual steps needed.

• CIFAR-100 / CelebA: Downloaded from HuggingFace and partitioned into federated splits. Pre-computed index files (dataset/*-splits.npz) are included in the repo. To build the dataset files, run:

  python -m dataset.build_dataset --dataset_name cifar100
  python -m dataset.build_dataset --dataset_name celeba

  Note: The disk usage for downloading & processing CelebA is ~28GB.

• SST-2: Pre-split data is already included in the repo under dataset/sst2-processed/. No download needed.

Running Experiments

Experiments are configured via src/constants.py. Default parameters are in init_parameters. To run a specific experiment, pass its index from var_params as an argument:

Predefined experiments without editing variables can be executed directly from the command line, as shown below. <index> specifies an experiment in the range [1-9] with 0 defaulting to a simple test.

Parallelism is controlled by the optional command line argument --num_proc; set to 0 or omit this argument for sequential execution. See the table below for recommended values for <num_proc>.

Commands:

python -m disagg_test --exp_index=<index> --num_proc=<num_proc>
python -m opa_test --exp_index=<index> --num_proc=<num_proc>
python -m light_secagg_test --exp_index=<index> --num_proc=<num_proc>
python -m secagg_plus_test --exp_index=<index> --num_proc=<num_proc>

For plain-text FL experiments, set USE_SECURITY=False in constants.py and use the disagg_test module.

Note: Wall clock times below refer to the full experiment required to produce the paper plots, measured on the reference hardware above.

Index	Description	Recommended num_proc	Wall Clock Time	Scripts to run
0	Quick test (N=100, M=1k)	16	<1m	disagg_test.py
1	Set M=N=10k and vary ρ (Figure 6)	16	~13.5h	{disagg|opa}_test.py
2	Sweep M, N with all protocols (Figure 7)	16	~25.7h	{disagg|opa|light_secagg|secagg_plus}_test.py
3	Set M=N=100k and vary ρ (Table 3)	4	>1d	{disagg|opa}_test.py
4	Set M=N=10k and vary γ,δ (Figure 8)	16	~23.2h	{disagg|opa}_test.py
5	Train NLP model with SST2 (Figure 9c-d)	2	~3.5h	{disagg|opa}_test.py for lora_r=16, disagg_test.py for lora_r=64
6	Train EfficientNet model with CELEBA (Figure 9e)	2	~4.3h	{disagg|opa}_test.py
7	Train CNN models on MNIST & CIFAR10 (Figure 9a-b)	16	~2.9h	{disagg|opa}_test.py
8	Train TinyNet model on CIFAR100 (Figure 9f)	4	~2.1h	{disagg|opa}_test.py
9	Experiment with stragglers on CIFAR10 (Figure 10)	16	~6.4h	{disagg|opa}_test.py

Example — run SST2 experiment using DisAgg:

python -m disagg_test --exp_index=5 --num_proc=2

Notes:

• When training with CIFAR10/100, the accuracy after 30 rounds is expected to be ~0.3/~0.5, due to the non-IID dataset splits used.
• Experiments create temporary files in the folder ./temp. This is created automatically at the beginning and its contents are deleted automatically once each experiment finishes. It is recommended to have at least 200-300GB free disk space for temporary files.

Experiment customization

Experiments can be customized by modifying the configuration parameters in src/constants.py. Edit init_parameters to change default values, or add new sweep configurations to var_params.

Citation

Note: This README corresponds to the artifact evaluation version of the paper.

Mehmood, H., Tatsis, G., Alexopoulos, D., Saravanan, K., Xu, J., Drosou, A., and Ozay, M. DisAgg: Distributed Aggregators for Efficient Secure Aggregation in Federated Learning. To appear in Proceedings of the Ninth Annual Conference on Machine Learning and Systems, MLSys 2026.