FedProTIP: Task-Agnostic Federated Continual Learning via Replay-Free Gradient Projection

Official implementation for our TMLR 2026 paper: FedProTIP

---

Overview

FedProTIP is a federated continual learning method designed for task-agnostic settings, where clients learn from sequentially arriving tasks without relying on explicit task identities during inference.

The method combines replay-free gradient projection with task identity prediction to mitigate catastrophic forgetting while preserving privacy constraints in federated learning.

This repository provides code to reproduce experiments on:

• 10-split CIFAR-100
• 6-split DomainNet
• ImageNet-R with 5, 10, and 20 task splits

For more details, please refer to our paper:

OpenReview: https://openreview.net/forum?id=GW4aw0fUKC

---

Installation

Install the required dependencies with:

pip install -r requirements.txt

---

Running Experiments

10-split CIFAR-100

python main.py --config configs/cifar100/fedprotip.json

---

6-split DomainNet

python main.py --config configs/domainnet/dom_fedprotip.json

---

ImageNet-R

5-split ImageNet-R

python main.py --config configs/imagenetr/imagenet-r_fedprotip.json --n_tasks 5 --increment 40

10-split ImageNet-R

python main.py --config configs/imagenetr/imagenet-r_fedprotip.json --n_tasks 10 --increment 20

20-split ImageNet-R

python main.py --config configs/imagenetr/imagenet-r_fedprotip.json --n_tasks 20 --increment 10

---

Code References

This repository builds on and adapts components from the following open-source implementations:

• LANDER
• GPM

We sincerely thank the authors for making their code publicly available.

---

Citation

If you find this repository useful, please cite our paper!

@article{cha2025task,
  title={Task-Agnostic Federated Continual Learning via Replay-Free Gradient Projection},
  author={Cha, Seohyeon and Chen, Huancheng and Vikalo, Haris},
  journal={arXiv preprint arXiv:2509.21606},
  year={2025}
}