Learnijoy
ResearchAI ResearchAI Engineering & DevTools

FedOGL Fights Catastrophic Forgetting in Federated Graph Learning.

Zekai Chen, Haodong Lu, Shihao Li, Weiwei Ji, Xunkai Li, Xun Wu, Yinlin Zhu, Rong-Hua Li· July 31, 2026 View original

Key takeaways

  • Catastrophic forgetting is a major challenge in federated open-world multimodal graph learning.
  • FedOGL is a framework designed to preserve semantic and structural memory.
  • It uses client-side replay/distillation and server-side prototype sharing.
  • FedOGL significantly reduces forgetting and improves performance.

Who benefits

HealthcareBFSITelecommunicationsIoTCybersecurity

Summary

This paper introduces FedOGL, a semantic-structural memory preservation framework designed to combat catastrophic forgetting in federated open-world multimodal graph learning. It enables clients to learn new classes while retaining old knowledge and rejecting unknown samples, addressing challenges like modality-semantic overwriting and topology-induced erosion.

Federated graph learning allows multiple clients to collaboratively train models on decentralized graph data without sharing raw information, which is crucial for privacy. However, in dynamic "open-world" scenarios where new data classes emerge, these systems face a significant challenge: catastrophic forgetting. This occurs when learning new information causes the model to lose previously acquired knowledge. This research highlights that in federated multimodal graphs, catastrophic forgetting is particularly complex, manifesting as the overwriting of old knowledge across different data modalities, erosion of structural information due to topology changes, and fragmentation of federated memory. To counter these issues, the authors propose FedOGL, a framework focused on semantic-structural memory preservation. FedOGL operates on both client and server sides. Clients preserve historical decision behavior through replay and task-start distillation, and protect graph-propagation memory by projecting it onto a globally shared structural basis. The server maintains and transfers compact category prototypes, facilitating knowledge sharing across clients without exposing sensitive raw graph data. Extensive experiments show FedOGL significantly reduces performance degradation from catastrophic forgetting by 42.67% compared to baselines, while maintaining or improving performance on downstream tasks.

Why it matters

Professionals building federated learning systems, especially those dealing with evolving data and multimodal graphs (e.g., in healthcare, finance, or IoT), can use FedOGL to create more robust, adaptable, and privacy-preserving models that don't forget past learnings.

How to implement this in your domain

  1. 1Assess existing federated learning pipelines for vulnerability to catastrophic forgetting, particularly with evolving data schemas.
  2. 2Investigate FedOGL's client-side memory preservation techniques (replay, distillation, projection) for integration into current client models.
  3. 3Explore implementing server-side compact category prototypes for efficient and private knowledge sharing.
  4. 4Pilot FedOGL on a specific federated learning task involving multimodal graph data to quantify improvements in knowledge retention.

Original post by Zekai Chen, Haodong Lu, Shihao Li, Weiwei Ji, Xunkai Li, Xun Wu, Yinlin Zhu, Rong-Hua Li

"arXiv:2607.27665v1 Announce Type: new Abstract: Federated graph learning enables collaborative training over decentralized graph data without sharing raw graph information. As such risks evolve, clients must learn emerging classes from private multimodal graph streams, retain his…"

View on X

Originally posted by Zekai Chen, Haodong Lu, Shihao Li, Weiwei Ji, Xunkai Li, Xun Wu, Yinlin Zhu, Rong-Hua Li on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

New Framework Improves Partial Multi-View Clustering Performance.

DAS-PMVC is a novel framework for partial multi-view clustering that addresses view asymmetry and irrelevant samples by leveraging dual alignment and structure enhancement. It uses anchor graph structure alignment, structure-enhanced feature learning, and a dual alignment strategy to achieve superior clustering performance on various datasets.

Shubin Ma, Liang Zhao, Chuanye He, Zhenjiao Liu, Liang Zou, Lin Yuanbo Wu, Yu ShaoJul 31, 2026
AI Engineering & DevToolsAI Research

Dual Teachers Improve Adversarial Robustness and Accuracy.

This work extends Information Bottleneck Distillation (IBD) by introducing a "clean teacher" alongside a robust teacher to improve the robustness/accuracy tradeoff against adversarial attacks. The proposed method transfers features from both teachers to a student model, achieving better clean accuracy while maintaining adversarial robustness, outperforming original IBD and competing with state-of-the-art approaches.

Vincent Ryusuke Takahashi, Yoshinari Takeishi, Jun'ichi Takeuchi, Kave SalamatianJul 31, 2026
AI Engineering & DevToolsAI Research

Dynamic Batch Sizes Improve Large Language Model Training Efficiency.

This paper proposes a new approach to deep learning dynamics, deriving joint scaling laws for loss based on both learning rate and batch size schedules. It introduces an optimal dynamic batch size schedule that consistently outperforms static batch size baselines, highlighting its importance for large language model training.

Jiaxiang Li, Zhiqi Bu, Shiyun XuJul 31, 2026