Research map / Federated learning and privacy / Personalized and heterogeneous FL
pFedMMA: Personalized Federated Fine-Tuning with Multi-Modal Adapter for Vision-Language Models
Sajjad Ghiasvand, Mahnoosh Alizadeh, Ramtin Pedarsani
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use pFedMMA instead of FedPGP | pFedMMA · CLIP datasets 16-shot (avg 7) | HM accuracy (%): 79.09 -> 84.15 | Improved |
| use pFedMMA instead of pFedMoAP on local classes | pFedMMA · CLIP datasets 16-shot (avg 7) | local accuracy (%): 97.89 -> 97.17 | Hurt |
| use pFedMMA on DomainNet instead of FedPGP | pFedMMA · DomainNet (Dirichlet 0.5) | avg accuracy (%): 24.90 -> 47.17 | Improved |
| use pFedMMA on Office-Caltech10 instead of FedPGP | pFedMMA · Office-Caltech10 (Dirichlet 0.5) | avg accuracy (%): 20.71 -> 21.33 | Improved |
| use pFedMMA instead of pFedMoAP at high label skew | pFedMMA · CIFAR-100 (beta=0.1) | accuracy (%): 80.29 -> 81.82 | Improved |
| use pFedMMA at mild heterogeneity (beta=10) vs FedPGP | pFedMMA · CIFAR-100 (beta=10) | accuracy (%): 73.90 -> 73.65 | No clear effect |
| keep up/down projections local, aggregate shared projection (vs all global) (ablation) | pFedMMA · DTD | accuracy (%): 61.10 -> 76.38 | Improved |
| keep up/down local and aggregate shared projection (vs shared local, up/down global) (ablation) | pFedMMA · DTD | accuracy (%): 62.19 -> 76.38 | Improved |
| use a single multi-modal shared projection instead of vision-only aggregation (ablation) | pFedMMA · CLIP datasets | HM accuracy (%): 79.31 -> 79.70 | Improved |
| raise scaling factor alpha from 0.0001 to 0.005 (ablation) | pFedMMA · CLIP datasets | HM accuracy (%): 7.62 -> 83.25 | Improved |
| raise scaling factor alpha from 0.005 to 0.01 (ablation) | pFedMMA · CLIP datasets | HM accuracy (%): 83.25 -> 82.03 | Hurt |
| insert adapters in last three layers instead of last layer only (ablation) | pFedMMA · CLIP datasets | HM accuracy (%): 83.57 -> 84.38 | Improved |
Later papers that compared with this method
2 later accepted papers report pFedMMA as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Beyond Description: Federated Adaptation via Semantic-Visual Prototype Alignment (ICML 2026 · better on 3 of 3 shared results)
- BLEND: Balancing Personalization vs. Generalization in Federated Vision–Language Models (NeurIPS 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Mixture of Experts Made Personalized: Federated Prompt Learning for Vision-Language Models (ICLR 2025 · similarity 0.91)
- BLEND: Balancing Personalization vs. Generalization in Federated Vision–Language Models (NeurIPS 2026 · similarity 0.91)
- Harmonizing Generalization and Personalization in Federated Prompt Learning (ICML 2024 · similarity 0.91)
- Federated Disentangled Tuning with Textual Prior Decoupling and Visual Dynamic Adaptation (ICML 2025 · similarity 0.90)
- F^3OCUS - Federated Finetuning of Vision-Language Foundation Models with Optimal Client Layer Updating Strategy via Multi-objective Meta-Heuristics (CVPR 2025 · similarity 0.89)
- Federated Learning from Vision-Language Foundation Models: Theoretical Analysis and Method (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
