Workshop Scope
The workshop will address both foundational and cutting-edge challenges in FL, integrating graph analytics with emerging trends like LLMs and generative AI. Unlike prior workshops (e.g., FedGraph@ICDM 2023 or FedKDD@KDD 2024), which centered on graph-structured data, this year’s focus expands to interdisciplinary synergies—from privacy-preserving LLM fine-tuning and multi-agent collaborative systems (e.g., autonomous vehicles) to FL-driven scientific workflows in climate modeling or biomedical research. We encourage submissions covering novel application domains, such as decentralized content generation and AI art platforms, while addressing systemic challenges like data heterogeneity, trust gaps, and scalability constraints.
(1) Scaling laws of FL facing increasingly larger and more heterogeneous data, models, and computation-or-communication resources. Essentially, we look for studies on how the effectiveness of existing or new FL algorithms changes by scaling. Specific topics include:
- Learning for larger language/vision models with more parameters, pre-training data, etc.;
- Learning with an increasing number of clients to billion scales;
- Learning from highly heterogeneous data distributions;
- Learning from heterogeneous hardware or computation capabilities with increasing gaps in hardware capabilities;
(2) Safety. Problems and solutions for the security, privacy, and social alignment of FL systems and the resultant models. Especially, when training large generative AI models, the potential risks and countermeasures in FL systems are welcome to discuss. Specific topics include:
- Privacy leakage during distributed training and inference;
- Risks for generative AI safety from the distributed data sources;
- Uncertainty in data collection with noisy labels or data;
- Backdoor attacks in training and inference;
- Data poisoning due to increasingly synthetic web data.
(3) Graph Analytics. Innovations to close the gap between centralized and decentralized graph analysis.
- Handling complex graph-structured data correlations and heterogeneity;
- Communication-efficient large-scale graph processing;
- Privacy preservation for both statistical and structural graph information;
- Evaluation of GNNs and FL algorithms in practical scenarios (e.g., knowledge graphs, e-commerce recommendations);
- Extending FL to complex graph variants (e.g., spatiotemporal networks, text-attributed graphs, manifolds);
- Federated graph algorithms beyond GNNs (e.g., spectral analysis, belief propagation);
- System optimization via graph mining principles.
(4) LLMs. Challenges in distributed training and alignment of large language models:
- Privacy-preserving pre-training/fine-tuning across decentralized data;
- Multi-tenant personalization and model alignment;
- Cross-modal FL (e.g., text-to-image with distributed corpora);
- Mitigating synthetic data poisoning in web-scale training;
- Efficiency optimization for trillion-parameter synchronization.
(5) AI-driven Science. Enabling secure cross-institutional collaboration in scientific domains:
- Multi-center biomedical studies (genomics, medicines);
- Distributed drug discovery with proprietary molecular data;
- Climate modeling via global sensor/web data federations;
- Physics simulations with privacy-sensitive experimental data;
Moreover, we aim to attract high-quality research on FL applications, evaluation methods, and algorithms. Open discussions on controversial FL topics (e.g., ethical trade-offs in collaborative AI) will be encouraged.
Submission Guidelines
Format:
We invite short technical papers - up to 5 pages including references and unlimited pages of appendix.
All manuscripts should be submitted in a single PDF file including all content, figures, tables, and references, following the new Standard ACM Conference Proceedings Template.
For LaTeX users: unzip acmart.zip, make, and use sample-sigconf.tex as a template; Additional information about formatting and style files is available online at ACM Proceedings Template.
Additionally, papers must be in the two-column format, with the recommended setting for Latex file: \documentclass[sigconf, anonymous, review]{acmart}.
Camera ready: Accepted papers can have up to 5 pages and unlimited pages of references and appendix. Please use the new acmart.cls when preparing submissions.
Submission: Papers should be submitted at the openreview website.
Review: All papers will be double-blinded and peer-reviewed by at least 2 reviewers.
Presentation: While all accepted papers will be presented with posters, high-quality accepted papers will also have the opportunity to participate in the oral/spotlight presentation and win our Best Paper Award(s). All accepted papers are expected to be presented in person. The workshop will not provide support for virtual talks or posters. We will also present accepted papers on our website.
According to the policy of the KDD conference, the accepted papers will NOT be included in proceedings or any form of publication.
Awards: The organizing committee will select best paper award(s) supported by our sponsors.
Important Dates
- Submission site open: April 1, 2025
- Paper submissions: May 26, 2025
- Paper notifications: June 22, 2025
- Early-bird registration due: TBD
- Workshop date: TBD