Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models

Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models Qianpiao Ma

Xiaozhu Song

Junlong Zhou∗

Nanjing University of Science and Technology Nanjing, Jiangsu, China maqianpiao@njust.edu.cn

Nanjing University of Science and Technology Nanjing, Jiangsu, China songxiaozhu@njust.edu.cn

Nanjing University of Science and Technology Nanjing, Jiangsu, China jlzhou@njust.edu.cn

Yue Zeng

Jianchun Liu

Huaqing Tu

Nanjing University of Science and Technology Nanjing, Jiangsu, China zengyue@njust.edu.cn

University of Science and Technology of China Hefei, Anhui, China jcliu17@ustc.edu.cn

Zhejiang Gongshang University Hangzhou, Zhejiang, China thq@mail.zjgsu.edu.cn

arXiv:2607.12583v1 [cs.DC] 14 Jul 2026

Abstract Federated fine-tuning (FedFT) enables adapting pre-trained large vision models (LVMs) on distributed, privacy-sensitive devices, while its practical deployment is hindered by three critical challenges: resource constraints, system heterogeneity, and non-IID data. While prior studies partially address these issues, e.g., by pre-training initial models on synthetic images to mitigate the adverse effects of non-IID data, or leveraging parameter-efficient finetuning (PEFT) methods like low-rank adaptation (LoRA) to reduce resource consumption, they remain inadequate and fragmented. Specifically, existing synthetic image generation methods fail to capture device-specific feature distributions, while current PEFT-based FedFT methods often undervalue weaker devices that may provide critical information. More importantly, the separate optimization of pre-training and FedFT neglects their inherent connection, lacking a holistic perspective to maximize training efficiency. To overcome these limitations, we propose FeDiSyn, a unified framework that holistically considers the interplay between pre-training and FedFT to minimize the overall LVM training time. Specifically, FeDiSyn introduces: (i) a scaling law for FedFT pre-training to determine the optimal number of synthetic images, balancing pre-training benefit against generation/pre-training cost, (ii) diffusion-based synthetic image generation that captures device-specific feature distributions for pre-training to tackle non-IID data, and (iii) a contributionaware LoRA configuration and bandwidth allocation algorithm for FedFT to ensure that informative devices are effectively utilized while addressing system heterogeneity. Experimental results on the real-world testbed demonstrate that FeDiSyn reduces completion ∗ Corresponding author.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. ICPP ’26, Singapore © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-XXXX-X/2018/06 https://doi.org/XXXXXXX.XXXXXXX

time by over 52.5% and communication cost by over 97.2%, while achieving comparable accuracy to state-of-the-art solutions.

CCS Concepts • Computing methodologies → Distributed artificial intelligence; • Human-centered computing → Mobile computing.

Keywords Federated fine-tuning, Synthetic image, Pre-training, Scaling law, Large vision models ACM Reference Format: Qianpiao Ma, Xiaozhu Song, Junlong Zhou, Yue Zeng, Jianchun Liu, and Huaqing Tu. 2026. Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models. In Proceedings of 55th International Conference on Parallel Processing (ICPP ’26). ACM, New York, NY, USA, 11 pages. https://doi.org/XXXXXXX.XXXXXXX

1

Introduction

Since the landmark success of AlexNet in 2012 [16], deep learning powered by large-scale labeled datasets has revolutionized the field of image classification. Initially, most vision models are trained via supervised learning on relatively small datasets. More recently, however, a two-stage paradigm has emerged for training large vision models (LVM): first pre-training on massive datasets, then fine-tuning on downstream task-specific data. In particular, pure transformer-based architectures such as the vision transformer (ViT) [5] have demonstrated more outstanding performance than the convolutional networks when pre-trained at scale. In parallel, the Internet of Things (IoT) [48] has grown explosively, generating massive amounts of data from edge devices that offers a rich source for training LVMs. However, centralized training on large-scale datasets is increasingly limited by privacy regulations (e.g., GDPR) and by bandwidth bottlenecks when accessing local data. This motivates the adoption of federated fine-tuning (FedFT) [46], a technique that integrates model fine-tuning into the federated learning (FL) framework [24]. In a typical FedFT setting, a parameter server first broadcasts a pre-trained global model to a set of distributed devices. Each device fine-tunes the model on its local data and sends the updated parameters back to the server for

ICPP ’26, September 28–October 1, 2026, Singapore

aggregation. This cycle of local fine-tuning and global aggregation repeats until the model converges or reaches a target accuracy. Challenges of FedFT for LVMs. Although FedFT offers a promising solution for fine-tuning LVMs leveraging distributed data, it still faces several practical challenges: (1) Resource Constraints. Modern LVMs like ViT and SwinT [21], usually contain billions of parameters, requiring significant computation/communication resources for fine-tuning/transmission. For example, ViT-L/16 requires roughly 3.1GB of traffic per communication and over 2,900 TFLOPs of computation. In contrast, device resources are typically limited [37]. For example, devices like the NVIDIA Jetson TX2 offer less than 2TFLOPs of computation capacity, while network bandwidth often falls below 100Mbps, rendering full-parameter FedFT infeasible on devices. (2) System Heterogeneity. Devices differ widely in hardware capabilities and network conditions. For example, the computational power of an iPhone 16 amounts to only about 12% of that of an RTX 3090 desktop GPU [18], while network condition varies significantly, ranging from 4G (∼ 100Mbps), 5G (∼ 500Mbps) to fiber connections (∼ 1Gbps). As a result, the time required for fine-tuning and transmitting parameters can differ substantially across devices. Faster devices may idle while waiting for slower ones, leading to a straggler problem [22] that reduces overall FedFT efficiency. (3) Non-IID Data. Local data on devices are often nonindependent and non-identically distributed [24], which involves two aspects. First, the label distribution varies across devices, e.g., some classes may be overrepresented or missing entirely on certain devices. Second, even for the same label, the feature distributions may differ due to variations in style or domain, e.g., cartoon vs. natural images [40]. Status Quo and Limitations. To address resource constraints in FedFT, some studies [8, 28, 29, 33, 34, 38, 39] have deployed the parameter-efficient fine-tuning (PEFT) methods (e.g., Adapter tuning, LoRA and Prompt tuning). PEFT inserts lightweight trainable modules into the LVM while keeping the pre-trained model frozen. Since these modules typically account for less than 5% of the pre-trained model’s parameters, PEFT significantly reduces the computation and communication overhead. However, most of these works [8, 28, 34, 38, 39] simply transplant centralized PEFT methods into distributed settings without accounting for system heterogeneity across devices. A few studies [29, 33] attempt to mitigate heterogeneity by assigning varying LoRA depths based on devices’ capabilities. This strategy, however, overlooks the potential importance of weaker devices in FedFT [2]: assigning them smaller LoRA depths can significantly degrade fine-tuning performance. Moreover, these FedFT approaches are typically initialized with off-the-shelf pre-trained models without synthetic data augmentation, indicating that convergence performance could be further improved [15]. Recent theoretical work [12] has shown that a carefully initialized pre-trained global model can mitigate the adverse effects of non-IID data and accelerate convergence in FedFT. However, models pre-trained on public datasets (e.g., ImageNet [4]) may not reflect the device-specific data distributions [3], making them suboptimal as initialization points for FedFT. Several studies [43, 45] have introduced diffusion models (DM) [32] to generate synthetic images that approximate the label distribution on each device. These synthetic images are then used to pre-train a global model, serving as

Ma et al.

the initialization for FedFT. However, these approaches rely solely on text-conditioned synthesis (e.g., labels or text embeddings) and may fail to capture the unique feature distribution of each device [47], leading to a style gap between synthetic and real data that diminishes the effectiveness of augmentation. Overall, none of these approaches adequately addresses the challenges outlined above, summarized in Table 1. More importantly, these approaches treat synthetic pre-training and FedFT as isolated stages, and ignore their synergy: insufficient images weaken pre-training effectiveness for FedFT, whereas excessive images introduce substantial computation overhead, thereby prolonging generation and pre-training time. Therefore, the key challenge lies in determining the optimal number of generated images, so as to balance pre-training benefit for FedFT against generation/pre-training cost. Overview of the Proposed Approach. To this end, we propose FeDiSyn, a unified framework that holistically considers pretraining and FedFT to minimize the overall completion time for LVM training, which is built upon “one law, two optimizations”. First, we introduce a scaling law for synthetic pre-training under FedFT setting, which captures the relationship among accuracy, total training time, and the synthetic data scaling factor, providing quantitative guidance for determining the optimal number of generated images. Second, for pre-training optimization, each device extracts features by encoding local images into latents, which are then uploaded to the server. By combining these latents with semantically enhanced labels via a DM, the server generates synthetic images to pre-train a global model. Since the synthetic images capture cross-device label and feature distributions, the pre-trained model yields a superior initialization for FedFT, effectively addressing both label and feature distribution skews. Third, for FedFT optimization, FeDiSyn introduces a contribution-aware LoRA configuration and bandwidth allocation algorithm. By configuring potentially important devices sufficient LoRA depth and balancing completion times via bandwidth allocation, the FedFT performance can be further improved while addressing system heterogeneity. Our main contributions are summarized as follows. • We propose FeDiSyn, a unified framework that holistically considers the close interplay between pre-training and FedFT to reduce the overall LVM training time. To our knowledge, this is the first work to explore the scaling law for synthetic pre-training under FedFT setting, and systematically optimize these two stages in a unified manner. • We develop a synthetic pre-training method that tackles nonIID data from both label and feature distribution perspectives. In addition, we design a contribution-aware LoRA configuration and bandwidth allocation algorithm to improve FedFT performance while addressing system heterogeneity. • We implement the proposed FeDiSyn framework and related algorithms on a real-world testbed. Experimental results on various models and datasets show that the proposed solution outperforms the state-of-the-art solutions.

2 Background and Motivation 2.1 Pre-Training Model Initialization for FedFT The Importance of Pre-Training and Initialization. As a central challenge in FedFT, non-IID data is typically addressed through

Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models

ICPP ’26, September 28–October 1, 2026, Singapore

Table 1: Comparison of FedFT methods for LVMs. Method PerAda [39] pFedSeq [28] CaFPT [8] FedPrefix [34] SFLF [38] FedRA [33] CAFF[29] FPS [3] GPT-FL [45] FGL [43] FeDiSyn (Ours)

Synthetic Image for Pre-training No No No No No No No Yes Yes Yes Yes

PEFT Method Adapter Adapter Prompt Prompt LoRA LoRA LoRA LoRA

Handling Resource Constraints ✓ ✓ ✓ ✓ ✓ ✓ ✓ × × × ✓

various aspects, such as device sampling [2] and aggregation mechanism [22, 23], while the impact of model initialization is often overlooked in most studies. Several recent works have highlighted the critical importance of pre-training for the initial model in FedFT. Nguyen et al. [26] observe that models initialized with pre-trained weights typically exhibit significantly lower training loss values at initialization compared to those initialized randomly. Karimireddy et al. [15] further demonstrate that this lower initial loss resulting from pre-training facilitates faster convergence during subsequent FedFT. Chen et al. [3] note that pre-training shapes a better conditioned loss surface and steers the model into a location closer to the optimal solution, enhancing the stability of model aggregation. These phenomena are theoretically supported by [12], which proves that pre-training effectively reduces sensitivity to certain hyperparameters within FedFT by diminishing misaligned filters at initialization. Synthetic Image for Pre-Training LVMs. Although pre-training initialization can enhance FedFT performance, models pre-trained on public datasets (e.g., ImageNet) show limited gains due to domain mismatch between the public data and the device-specific data. For example, in chest X-ray analysis, an ImageNet pre-trained model achieves only 0.2% accuracy improvement over random initialization, whereas a domain-relevant synthetic image pre-trained model yields a 5.9% gain [3]. To address this, GPT-FL [45] and FGL [43] upload labels or text descriptions of device datasets to the server for synthetic image generation. However, these approaches rely solely on text-conditioned generation, which fails to capture devicespecific feature distributions. As a result, the synthetic images may not align well with real local data, limiting their improvement for FedFT. In contrast, FeDiSyn uploads latents encoding local data features to generate synthetic images, as elaborated in Section 3. For example, generated images for Caltech-101 [7] in Fig. 1 show that FeDiSyn can obtain synthetic images that better align with the feature distributions of raw data compared to GPT-FL and FGL.

2.2

Scaling Law of Synthetic-Image Pre-Training for FedFT

Following the groundbreaking success of large-scale models, scaling laws have emerged as a critical framework for quantifying the power-law relationships between model/data scale and performance [14]. Prior studies have demonstrated that pre-training data

Handling System Heterogeneity × × × × × ✓ ✓ × × × ✓

Image

Chair

Handling Label Distribution Skew × × × × × × × ✓ ✓ ✓ ✓

Cup

Ewer

Handling Feature Distribution Skew × × × × × × × × × × ✓

Ferry

Ibis

Raw

Fig. 1: Comparison of FeDiSyn and other synthetic image generation methods on Caltech-101. volume has a significant impact on model performance [27, 42], which has motivated recent extensions of these principles to synthetic data, exploring how generative data scale influences convergence [6, 13]. However, these investigations focus exclusively on centralized training paradigms and are therefore not directly applicable to the FedFT setting, where the capabilities of servers for pre-training and edge devices for fine-tuning jointly determine the overall training time. As illustrated in Fig. 2 on Caltech-101, the synthetic scaling factor 𝜆 introduces a trade-off in FedFT: increasing 𝜆 accelerates convergence but incurs substantial overhead in image generation and pre-training. Therefore, the total completion time follows a U-shaped curve, which, in this example, attains its minimum near 𝜆 = 0.5. To provide guidance for synthetic pre-training in the FedFT setting, we propose a novel scaling law that quantifies the relationship between synthetic volume, target accuracy, and total training time, enabling the analytical determination of the optimal scaling factor 𝜆, as elaborated in Section 5.

2.3

Federated Fine-Tuning with LoRA for LVMs

LoRA [11] fine-tuning is a specific PEFT approach via low-rank matrix decomposition. In LoRA-based FedFT, each device only finetunes and uploads the lightweight inserted LoRA layers while keeping the pre-trained model frozen. The server aggregates the uploaded LoRA parameters and then distributes them back to all

ICPP ’26, September 28–October 1, 2026, Singapore



① Synthetic Image Generation

*HQHUDWLRQ 3UHWUDLQLQJ )HG)7



5LFKSRZHUIXO 5LFKZHDN 3RRUSRZHUIXO 3RRUZHDN

 

      

 

/R5$    GHSWK  V

         

6FDOLQJIDFWRU





)LQHWXQLQ



Encode

$FFXUDF\

7RWDOWLPHî2V



Ma et al.

VAE

Pre-Trained Model Parameters

LoRA Parameters

LLM Prompt

Feature Extraction



JWLPHî

Device 1

③Federated Fine-Tuning Label

CLIP Text Encode

V ...

Fig. 2: The effect of different Fig. 3: The tradeoff between 𝜆 on total time to reach 70% heterogeneity and perforaccuracy on Caltech-101. mance.

Text Embedding

Scheduler

② Model Pre-Training Global Model

Add Noise

Latent Variable

Diffusion Model Predict

devices. Most existing LoRA-based FedFT approaches for LVMs [33, 38] insert LoRA layers into all transformer layers, causing significant computation overhead. To reduce this burden, a few studies [19, 29] insert LoRA only to a subset of layers, enabling more efficient fine-tuning on resource-constrained devices. Futhermore, these works typically address system heterogeneity by assigning LoRA depths based solely on device capabilities: powerful devices receive deeper LoRA layers, while resource-constrained devices get smaller ones. However, weak devices may play a critical role in FedFT, and assigning them shallow LoRA depths can significantly degrade performance. We conduct a case study with ViT-B/16 fine-tuned on CIFAR-10 [17] in Fig. 3, which compares accuracy and fine-tuning time across four device types: (i) richdata/powerful-capability, (ii) rich-data/weak-capability, (iii) poordata/powerful-capability, and (iv) poor-data/weak-capability. As shown, powerful but data-scarce devices (red line) gain little benefit from deeper LoRA depths, whereas data-rich devices with weakcapability (green line) achieve accuracy comparable to powerful ones (blue line) when given a deeper LoRA, albeit with longer finetuning time. These observations highlight a trade-off: It is worth setting a larger LoRA depth for resource-constrained yet critical devices, but these devices may become stragglers that delay the overall FedFT process. To this end, we propose a contribution-aware LoRA configuration and bandwidth allocation strategy in FeDiSyn. LoRA depth for each device is determined by its expected performance gain, while bandwidth is allocated to balance device completion times, effectively addressing system heterogeneity. Details are elaborated in Section 5.

3 System Design of FeDiSyn 3.1 Preliminaries We consider a 𝐾-class classification problem performed through FedFT with label space C = {𝑐 1, 𝑐 2, ..., 𝑐 𝐾 }. The system contains a centralized server with relatively abundant computational resources and a set of devices V = {𝑣 1, 𝑣 2, ..., 𝑣 𝑁 }. Each device 𝑣𝑖 owns a local dataset D𝑖 , with size 𝐷𝑖 = |D𝑖 |. The global dataset across all Í devices is represented as D, with size 𝐷 = |D| = 𝑣𝑖 ∈ V 𝐷𝑖 . The data labeled as 𝑐𝑘 on 𝑣𝑖 is denoted as D𝑖𝑘 with size 𝐷𝑖𝑘 = |D𝑖𝑘 |, and Í it holds that 𝑐𝑘 ∈ C 𝐷𝑖𝑘 = 𝐷𝑖 . The loss function of 𝑣𝑖 is defined as 𝑓𝑖 (w) ≜ 𝐷1𝑖

∑︁ 𝜉 ∈ D𝑖

𝑓𝑖 (w; 𝜉),

(1)

Synthetic Images

Server

Label

Images

Loss Function

Fig. 4: Overview of FeDiSyn. where w is the parameter vector, and 𝑓𝑖 (w; 𝜉) is the loss over sample 𝜉 in D𝑖 . The global loss function on all the distributed datasets is ∑︁ 𝐷𝑖 (2) 𝐹 (w) ≜ 𝐷 𝑓𝑖 (w). 𝑣𝑖 ∈ V

The learning problem is to find the optimal parameter vector w∗ so as to minimize 𝐹 (w), i.e., w∗ = arg minw 𝐹 (w).

3.2

System Overview

As shown in Fig. 4, the process of FeDiSyn consists of three steps. 1 Synthetic image generation: Devices encode private images ○ into low-dimensional latents using a lightweight variational autoencoder (VAE). These latents are clustered to obtain class-specific centroids and a global average representation, which are then sent to the server with their labels. The server uses an large language model (LLM) to expand labels into diverse text prompts, encodes them via CLIP’s [30] text encoder, and feeds the resulting embeddings along with the uploaded latents into a DM (e.g., Stable Diffusion) to generate synthetic images, where latents govern stylistic features (e.g., lighting, texture) and text embeddings condition se2 Model pre-training: The mantics (e.g., class-specific details). ○ server pre-trains a global model over synthetic images, yielding an initial model that captures cross-device features for subsequent 3 Federated fine-tuning: The initialized model is sent to FedFT. ○ devices for LoRA-based fine-tuning under the FedFT framework. Specifically, each device keeps the pre-trained model frozen and updates only LoRA parameters, which are aggregated by the server and redistributed until convergence. The following sections detail each step.

3.3

Synthetic Image Generation

Data Feature Extraction. To align synthetic images with devicespecific styles while preserving privacy, each device 𝑣𝑖 extracts the features of its raw images locally in the latent space. Specifically, 𝑘 ∈ D 𝑘 of class 𝑐 on device 𝑣 is encoded by a each sample D𝑖,𝑗 𝑖 𝑘 𝑖 𝑘 = E (D 𝑘 ). Subsequently, for the same VAE E to obtain latents 𝑒𝑖,𝑗 𝑖,𝑗 𝑘 , ∀𝑗 into 𝑍 cluster centroids class 𝑐𝑘 , device 𝑣𝑖 clusters the latents 𝑒𝑖,𝑗 𝑘 e𝑖,𝑧 , 𝑧 ∈ [1, 𝑍 ], and calculate their average representation as e𝑘𝑖 . e𝑘𝑖,𝑧 and e𝑘𝑖 capture the style characteristics of the device distribution

Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models

[41]. Then the device uses a scheduler to add noise to the encoded latents at a randomly chosen diffusion timestep 𝑠 ∈ {0, ..., 𝑆 } [31] as √ √ 𝑘 ē𝑘,𝑠 𝑖,𝑧 = 𝛼𝑠 e𝑖,𝑧 + 1 − 𝛼𝑠 𝜖, 𝜖 ∼ N (0, 𝐼 ), 𝑧 ∈ [1, 𝑍 ], (3) √ 𝑘 √ ē𝑘,𝑠 𝑖 = 𝛼𝑠 e𝑖 + 1 − 𝛼𝑠 𝜖, 𝜖 ∼ N (0, 𝐼 ), where 𝛼𝑠 controls noise intensity at timestep 𝑠. The device retains the raw data locally and only uploads the cluster centroids ē𝑘,𝑆 𝑖,𝑧 , the average representation ē𝑘,𝑆 , and its class label 𝑐 to the server. 𝑘 𝑖 Label Collection and Prompt Engineering. The server collects the data volume 𝐷𝑖𝑘 labeled as 𝑐𝑘 of each device 𝑣𝑖 and calculates the total amount of data labeled as 𝑐𝑘 across all devices Í by 𝐷 𝑘 = 𝑣𝑖 ∈ V 𝐷𝑖𝑘 . To enrich semantic variety, a deployed LLM expands each label into multiple descriptive prompts. It has been verified that language enhancement is an effective way for enlarging synthetic image diversity for pre-training [9]. For example, for the label “dog”, the prompt can be semantically expanded as follows: Q: Write a sentence using the word “dog” to describe the characteristics and actions of the object. A: A golden retriever playing on the grass. These prompts are encoded into text embeddings via CLIP’s text encoder, ensuring semantic alignment between text and image latent spaces. To preserve the real data distribution, the synthetic image size for class 𝑐𝑘 is set as 𝐷ˆ 𝑘 = 𝜆𝐷 𝑘 . 𝜆 is the scaling factor, which is determined according to our proposed scaling law in Section 5. Synthetic Image Generated by Diffusion Model. On the server, for each class 𝑐𝑘 , the received cluster centroids ē𝑘,𝑆 𝑖,𝑧 and ¯𝑘,𝑆 = 𝜌 ē𝑘,𝑆 +(1−𝜌)ē𝑘,𝑆 , average representation ē𝑘,𝑆 are weighted as 𝐸 𝑖,𝑧 𝑖,𝑧 𝑖 𝑖 where 𝜌 is a factor with a value between 0 and 1. Next, each latent 𝑘,𝑆 is paired with a text embedding and fed into a latent DM, where 𝐸¯𝑖,𝑧 latents guide stylistic consistency and text embeddings enforce data diversity. At each reverse step, a PNDM scheduler [20] is employed 𝑘,𝑆 for sampling, iteratively denoising 𝐸¯𝑖,𝑧 over 𝑆 steps via √ 𝛼 (𝛼 𝑠 −𝛿 −𝛼 𝑠 ) 𝑘,𝑆 𝑘,𝑆  𝜖𝑠 , 𝜙 (𝐸¯𝑖,𝑧 , 𝜖𝑠 , 𝑠, 𝑠 − 𝛿) = √𝛼𝑠 −𝛿 𝐸¯𝑖,𝑧 − √ √ √ 𝑠

𝛼𝑠

(1−𝛼 𝑠 −𝛿 )𝛼 𝑠 +

(1−𝛼 𝑠 )𝛼 𝑠 −𝛿

(4) where 𝛿 is the step interval, and the gradient term 𝜖𝑠 is computed using numerical methods such as the Runge-Kutta method or the 𝑘,𝑆 linear multi-step method. Ultimately, all 𝐸¯𝑖,𝑧 , ∀𝑖, 𝑘 undergo a certain number of denoising process to produce synthetic images D̂ with Í Í size 𝐷ˆ = 𝑐𝑘 ∈ C 𝐷ˆ 𝑘 = 𝜆 𝑐𝑘 ∈ C 𝐷𝑘 at server side. Remark 1 (Privacy Analysis). Transmitting latent representations is a widely used strategy for data augmentation in FL. Considering the limited amount of latent information uploaded by each device, our approach introduces minimal privacy leakage. Furthermore, similar to [41], we generate images using DM from noise-added features. As a result, even when the original images contain sensitive content (e.g., text or faces), the generated images exhibit entirely different semantic details. Therefore, despite potential visual similarities, they do not reveal or compromise the underlying private information.

3.4

Model Pre-Training over Synthetic Image

With the generated synthetic images D̂ on the server, a global model is pre-trained in a standard supervised fashion. Similar to Eq.

ICPP ’26, September 28–October 1, 2026, Singapore

Server

Device

① Global Model Initialization

③ Local LoRA Assignment

② LoRA Configuration

⑤ LoRA Aggregation

④ Local Model Fine-Tuning

Fig. 5: Federated fine-tuning with LoRA. (2), we denote 𝐹ˆ (w) as the loss function over the synthetic images D̂ for model parameters w. Starting from an initial parameter vector ŵ0 , the server performs 𝑇ˆ iterations of model pre-training by ŵ𝑡 = ŵ𝑡 −1 − 𝜂∇𝐹ˆ ( ŵ𝑡 −1 ),

(5)

where 𝜂 is the learning rate, and ∇ is the gradient operator. The ˆ global model after 𝑇ˆ iterations of pre-training is denoted by w = ŵ𝑇 , which acts as the initial model for the subsequent FedFT process.

3.5

Federated Fine-Tuning with LoRA

After the pre-training model w is obtained, FeDiSyn performs FedFT with LoRA for the LVM through the five phases shown in Fig. 5. 1 Global Model Initialization. The server first injects the ○ LoRA layers into the pre-trained model w, i.e., embedding initial LoRA matrices w e 0 = {e w𝑙0, ∀𝑙 ∈ [0, 𝐿 − 1]} within 𝐿 transformer layers of the neural network. Then the embedded pre-trained model w0 = {w, w e 0 } is distributed to all devices in V. 2 ○ LoRA Configuration. Prior studies [19, 44] show that inserting LoRA into deeper layers yields better fine-tuning performance than shallow layers. Therefore, we define LoRA depth as the number of layers with LoRA modules, counted from the deepest output layer upward. At each round 𝑡, the server assigns each device 𝑣𝑖 a depth 𝑥𝑖𝑡 based on its heterogeneous resources, thus only the last 𝑥𝑖𝑡 layers 𝐿𝑖𝑡 = {𝑙 |𝐿 − 𝑥𝑖𝑡 ≤ 𝑙 < 𝐿} are fine-tuned. The LoRA configuration strategy for all devices at round 𝑡 is denoted as X𝑡 = {𝑥𝑖𝑡 |𝑣𝑖 ∈ V}, which will be described in Section 5. 3 Local LoRA Assignment. At each round 𝑡, the server dis○ tributes the configuration information X𝑡 and the global LoRA parameters w e 𝑡 = {e w𝑙𝑡 |∀𝑙 ∈ [0, 𝐿 − 1]} to all devices in V. The 𝑙-th LoRA layer of device 𝑣𝑖 ’s local model is assigned as 𝑡 w e𝑖,𝑙 =w e𝑙𝑡 .

(6)

4 Local Model Fine-Tuning. After the local LoRA configura○ tion, 𝑣𝑖 fine-tunes the local model w𝑖𝑡 = {w, w e𝑖𝑡 } on its local dataset 𝑡 D𝑖 . The update of each LoRA layer w e𝑖,𝑙 at round 𝑡 is expressed as ( 𝑡 − 𝜂∇𝑓 (e 𝑡 𝑡 w e𝑖,𝑙 𝑖 w𝑖,𝑙 ), 𝐿 − 𝑥𝑖 ≤ 𝑙 < 𝐿 𝑡 +1 w e𝑖,𝑙 = . (7) 𝑡 w e𝑖,𝑙 , otherwise 5 LoRA Aggregation. After local fine-tuning, each device 𝑣𝑖 ○ 𝑡 +1 , 𝐿 − 𝑥 𝑡 ≤ 𝑙 < 𝐿 to uploads the fine-tuned LoRA parameters w e𝑖,𝑙 𝑖 the server. Upon receiving the updated LoRA parameters from all devices, for each layer 𝑙, the server aggregates them as Í 𝑡 𝑡 +1 e𝑖,𝑙 𝑣𝑖 ∈ V [𝑙 ≥ 𝐿 − 𝑥𝑖 ]𝐷𝑖 w 𝑡 +1 , (8) w e𝑙 = Í 𝑡 𝑣𝑖 ∈ V [𝑙 ≥ 𝐿 − 𝑥𝑖 ]𝐷𝑖

where [·] is the Iverson notation. As a result, the global LoRA parameters are updated as w e 𝑡 +1 = {e w𝑙𝑡 +1 |∀𝑙 ∈ [0, 𝐿 − 1]}, and the server configures LoRA for the next round 𝑡 +1. The process repeats until the model converges or reaches satisfactory accuracy.

ICPP ’26, September 28–October 1, 2026, Singapore

4

Convergence Analysis

We make the following assumptions on the loss functions 𝐹 (w) for convergence analysis, which are widely used in prior works [37]. Assumption 1 (Lipschitz). The loss function 𝐹 has a L-Lispschitzian gradient for ∀w1, w2 , i.e., ∥∇𝐹 (w1 ) − ∇𝐹 (w2 )∥ ≤ L ∥w1 − w2 ∥. Assumption 2 (Gradient Bounded). The expected squared norm of gradients is uniformly bounded for ∀w𝑖𝑡 , i.e., E∥∇𝐹 (w𝑖𝑡 )∥ 2 ≤ 𝐺 2 . For LoRA fine-tuning, each LoRA matrix is further decomposed as w e𝑖𝑡 = B𝑖𝑡 A𝑖𝑡 , ∀𝑖, 𝑡, where B𝑖𝑡 and A𝑖𝑡 are low-rank matrices. Assumption 3 (LoRA Bounded). There exist constants 𝐶 A , 𝐶 B > 0, such that ∥A𝑖𝑡 ∥ 2 ≤ 𝐶 A , ∥B𝑖𝑡 ∥ 2 ≤ 𝐶 B . Based on the above assumptions, we derive the convergence bound of FeDiSyn by the following theorem. Theorem 1 (Convergence of FeDiSyn). After 𝑇 rounds of finetuning in FeDiSyn, the expectation of the gradient of the LoRA matrices satisfies
1 ∑︁𝑇 ∑︁ E∥∇𝐹 (A𝑖𝑡 )∥ 2 + E∥∇𝐹 (B𝑖𝑡 )∥ 2 𝑡 =1 𝑣𝑖 ∈ V 𝑇𝑁 E[𝐹 (w0 )] − E[𝐹 (w𝑖∗ )] 1 ∑︁ ≤ + 𝜅, (9) 𝑣𝑖 ∈ V 𝑁 𝜂𝑇 where w𝑖∗ denotes the optimal model on 𝑣𝑖 , 𝜅 is a constant satisfying 𝐺2 2 2 2 2 2 4 2 2 [𝐺 3𝐶 A 𝐶 B + 3L 2 (𝐺 𝐶 A + 𝐺 𝐶 B + 𝜂 𝐺 𝐶 A 𝐶 B )]𝜂 + 2 ≤ 𝜅. Due to limited space, we omit the proof of Theorem 1 here. Remark 2 (LoRA-based FedFT Convergence). It is noticed that, unlike prior convergence analyses for FedFT [19, 38], which typically study the convergence behavior only at the level of the overall model parameter w, our analysis is developed specifically for LoRA-based FedFT and explicitly characterizes the convergence of the two low-rank matrices A and B. As a result, our convergence bound provides a fine-grained theoretical understanding of the training in LoRA-based FedFT, and better reveals how the low-rank decomposition influences convergence.

5

Problem Formulation and Algorithm Description 5.1 Problem Formulation The total completion time of FeDiSyn includes the image generation, pre-training, and FedFT time. Let ℎˆ denote the time to generate one image on server, then the total image generation time is expressed ˆ Next, the server pre-trains LVM on synthetic as 𝐻 gene = 𝐷ˆ ℎˆ = 𝜆𝐷 ℎ. data for 𝑇ˆ rounds, and the pre-training time can be expressed as 𝐻 pre = 𝜆𝐷 ℎ˜𝑇ˆ , where ℎ˜ is the single round training time for one image on server. For FedFT, although deeper LoRA layers improve its performance, practical deployments are constrained by resource limitations such as computation, communication, and memory. Let ℎ𝑖 denote the time for data loading and performing a full forward pass through the LVM on 𝑣𝑖 per sample, and ℎ𝑖 denote the additional time for backpropagating through per LoRA layer. Since the computation capacity of 𝑣𝑖 is fixed, both ℎ𝑖 and ℎ𝑖 can be measured in advance. Therefore, the fine-tuning time on device 𝑣𝑖 at round 𝑡 is given by

Ma et al. 𝑡,comp

𝐻𝑖 = 𝐷𝑖 (ℎ𝑖 + 𝑥𝑖𝑡 ℎ𝑖 ). The transmission time primarily arises from uploading LoRA parameters via FDMA, which allows simultaneous transmissions from multiple devices [38]. Let 𝐵𝑖𝑡 denote the bandwidth allocated to 𝑣𝑖 at round 𝑡, then the uplink transmis𝑝 𝑖 ℎ𝑡

sion rate 𝑟𝑖𝑡 = 𝐵𝑖𝑡 log(1 + 𝜎 2𝑖 ) is modeled by the Shannon capacity, where 𝑝𝑖 is the transmission power, ℎ𝑖𝑡 is the channel gain, and 𝜎 2 is the noise power. Therefore, the transmission time for 𝑣𝑖 at round 𝜏𝑥 𝑡

𝑡 is 𝐻𝑖𝑡,comm = 𝑟 𝑡𝑖 , where 𝜏 is the parameter size per LoRA layer. 𝑖

The duration of round 𝑡 is determined by the slowest device, i.e., 𝑡,comp

𝐻 𝑡 = max𝑣𝑖 ∈ V {𝐻𝑖

+ 𝐻𝑖𝑡,comm }.

(10)

For memory, let 𝑚 denote the fixed memory to load the LVM and store forward activations, and 𝑚 represent the extra memory incurred by gradient storage per LoRA layer during backpropagation. Then, the total memory overhead for fine-tuning on 𝑣𝑖 at round 𝑡 is 𝑚𝑖𝑡 = 𝑚 + 𝑥𝑖𝑡 𝑚.

(11)

Í Let the total completion time 𝐻 = 𝐻 gene + 𝐻 pre + 𝑇𝑡=1 𝐻 𝑡 , we formulate the problem as follows: (P1) : min𝜆,X𝑡 ,B𝑡 𝐻 s.t.

𝐹 (w𝑇 ) − 𝐹 (w∗ ) ≤ 𝜖 ∑︁ 𝐵𝑖𝑡 ≤ 𝐵

𝑣𝑖 ∈ V 𝑚𝑖𝑡 ≤ 𝑀𝑖

(12a) (12b) ∀𝑡 ∈ [0,𝑇 ]

(12c)

∀𝑖 ∈ [1, 𝑁 ]

(12d)

Constraint (12b) represents the global model converges after 𝑇 rounds, where 𝜖 is the convergence threshold. Constraint (12c) ensures the total allocated bandwidth does not exceed the budget 𝐵 in each round. Constraint (12d) ensures each device 𝑣𝑖 ’s memory usage does not exceed its available memory 𝑀𝑖 . Our objective is to decide the data scaling factor 𝜆, the LoRA depth configuration X𝑡 = {𝑥𝑖𝑡 |𝑣𝑖 ∈ V} and bandwidth allocation B𝑡 = {𝐵𝑖𝑡 |𝑣𝑖 ∈ V} for each round, minimizing the total completion time, i.e., min𝜆,X𝑡 ,B𝑡 𝐻 .

5.2

Algorithm Description

In this section, we introduce the algorithm description for our FeDiSyn to solve problem P1, as described in Alg. 1. 1 Determining Scaling Factor. The key challenge of solving ○ problem P1 is determining the scaling factor 𝜆 to balance the time required for synthetic image generation/pre-training and FedFT. From Section 5.1, the total time for generation and pre-training is 𝐻 gene + 𝐻 pre = 𝜆𝐷 ℎˆ + 𝜆𝐷 ℎ˜𝑇ˆ = 𝜆(ℎˆ + ℎ˜𝑇ˆ )𝐷 = 𝜆ℎ 0 , where ℎ 0 ≜ ˆ ℎ˜𝑇ˆ )𝐷 is defined as the unit time for generation and pre-training. (ℎ+ Conversely, empirical analysis across multiple datasets suggests that the subsequent FedFT time is negatively correlated with 𝜆 and grows exponentially with the target accuracy 𝐴𝑐𝑐. To quantify this relationship, we propose the following scaling law to describe the total completion time as 𝐻 (𝜆, 𝐴𝑐𝑐) = ℎ 0 𝜆 + 𝜅1 𝜆1𝜅2 𝑒 𝜅3 𝐴𝑐𝑐 + 𝜅 4,

(13)

where ℎ 0 𝜆 represents the linear growth of generation and pretraining time, and the remaining terms capture the diminishing returns of synthetic data on FedFT acceleration. The architecturespecific coefficients 𝜅 1 -𝜅 4 , which depend on the dataset, model and device capacity, can be estimated in a short calibration phase using a set of small-scale pilot runs, detailed in Section 6.2. To

Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models

Algorithm 1 Diffusion Model-Driven Synthetic Image Pre-Training for Federated Fine-Tuning (FeDiSyn) Input: Data volume 𝐷𝑖 , available memory 𝑀𝑖 , bandwidth budget 𝐵 1: /* Step 1. Determining Scaling Factor.*/ 2: Conduct a set of small-scale experiments to measure 𝐻 (𝜆, 𝐴𝑐𝑐) 3: Fit coefficients 𝜅 1 -𝜅 4 using nonlinear least squares method

√︃

𝜅 2 +1 𝜅 2 𝑒 𝜅 3 𝐴𝑐𝑐

4: Obtain optimal scaling factor 𝜆 ∗ =

ℎ0𝜅1

5: /* Step 2. Generating Synthetic Images and Pre-training.*/ 6: Obtain the centroids e𝑘𝑖,𝑧 and average representation e𝑘𝑖 on 𝑣 𝑖 𝑘,𝑠

𝑘,𝑠

7: Devices upload noisy latents ē𝑖,𝑧 , ē𝑖

and labels of each 𝑐𝑘

8: Server expands the labels C into prompts 9: Server performs image generation with 𝜆 ∗ yields D̂ 10: Pre-train LVM over D̂ and distribute LoRA-inserted models 11: /* Step 3. Federated Fine-tuning.*/ 12: for each 𝑡 ∈ 𝑇 do

/* Contribution-aware LoRA configuration and bandwidth allocation algorithm (CLCBA).*/ 14: for each 𝑣𝑖 ∈ V do 15: 𝑥𝑖𝑡 ← 1 , 𝑥𝑖max ← min (⌊(𝑀𝑖 − 𝑚)/𝑚}⌋ , 𝐿) 16: 𝑅𝑡max ← 0 17: while true do 18: for each 𝑣𝑖 ∈ V do 19: if 𝑥𝑖𝑡 < 𝑥𝑖max then 20: 𝑥𝑖𝑡 ← 𝑥𝑖𝑡 + 1 Obtain B𝑡 under X𝑡 by water-filling based bandwidth 21: allocation sub-algorithm (WFBA, Alg. 2) 22: Calculate completion time 𝐻 𝑡 by Eq. (10) 23: Calculate performance gain 𝑅𝑡 by Eq. (14) 24: Δ𝑅𝑖𝑡 ← 𝑅𝑡 − 𝑅𝑡max , 𝑥𝑖𝑡 ← 𝑥𝑖𝑡 − 1 25: if ∃Δ𝑅𝑖𝑡 > 0 then 26: 𝑖 ← argmax𝑣𝑖 ∈ V Δ𝑅𝑖𝑡 , 𝑅𝑡max ← 𝑅𝑡max + Δ𝑅𝑖𝑡 , 𝑥𝑖𝑡 ← 𝑥𝑖𝑡 + 1 27: else 28: break 29: Obtain B𝑡 under X𝑡 by Alg. 2 30: After LoRA assignment, devices perform local fine-tuning 31: Server aggregates the LoRA matrices uploaded by devices 32: return The fine-tuned model w𝑇

13:

find the optimal scaling factor 𝜆 ∗ that minimizes the total training time, we take the derivative of Eq. (13) with respect to 𝜆 as 𝐻𝜆′ = ℎ 0 − 𝜅 𝜆𝜅𝜅22 +1 𝑒 𝜅3 𝐴𝑐𝑐 . Taking 𝐻𝜆′ = 0 yields the optimal scaling factor 1 √︃ 𝜅 2 +1 𝜅 2 𝑒 𝜅 3 𝐴𝑐𝑐 ∗ 𝜆 = ℎ 0 𝜅 1 . We derive the following key insights regarding 𝜆∗ . Remark 3 (Impact of Server Capability). As the server’s computational capacity increases (ℎ 0 → 0), the optimal scaling factor 𝜆 ∗ grows. When the server possesses plenty of computational capacity, it can rapidly generate a vast synthetic images. As a result, a larger 𝜆 accelerates FedFT without incurring a significant generation time. Remark 4 (Impact of Target Accuracy). The presence of 𝐴𝑐𝑐 in 𝜅3

𝐴𝑐𝑐

the exponent, 𝜆 ∗ ∝ 𝑒 𝜅2 +1 , reveals that the demand for synthetic data grows exponentially with the target accuracy. For low-accuracy

ICPP ’26, September 28–October 1, 2026, Singapore

Algorithm 2 Water-Filling based Bandwidth Allocation (WFBA) Input: Bandwidth budget 𝐵, depth configuration X𝑡 , transmission power 𝑝𝑖 , and channel gain ℎ𝑖𝑡 Output: Optimal bandwidth allocation B𝑡 1: for 𝑖 ← 1 to 𝑁 do 𝑡,comp 2: Calculate 𝐻𝑖 ← 𝐷𝑖 (ℎ𝑖 + 𝑥𝑖𝑡 ℎ𝑖 ) 𝑝 𝑖 ℎ𝑡

𝛾𝑖𝑡 ← 𝜏𝑥𝑖𝑡 /log(1 + 𝜎 2𝑖 ) 𝑡,comp 𝑡,comp ← max𝑖 𝐻𝑖 4: 𝐻 max 𝑡,comp 5: 𝑞 ← 0, H𝑞𝑡 ← 2𝐻 max 6: while |Δ𝐻 𝑡 | > 𝜁 do Í𝑁 𝛾𝑖𝑡 7: Φ(H𝑞𝑡 ) = 𝑖=1 𝑡,comp − 𝐵 𝑡 3:

9:

H𝑞 −𝐻𝑖 Φ( H𝑞𝑡 ) 𝑡 𝑡 H𝑞+1 = H𝑞 − Φ′ ( H𝑡 ) 𝑞 𝑡 Δ𝐻 𝑡 ← H𝑞+1 − H𝑞𝑡

10:

𝑞 ←𝑞+1

8:

11: for 𝑖 ← 1 to 𝑁 do 𝑡,comp

𝐵𝑖𝑡 ← 𝛾𝑖𝑡 /(H𝑞𝑡 − 𝐻𝑖 13: return B 𝑡 12:

)

requirements, the overhead of large-scale synthesis outweighs its benefits. However, for high-accuracy requirements, the FedFT time increases so sharply that a massive synthetic pre-training for model initialization becomes essential to reduce the total training time. 2 Generating Synthetic Images and Pre-Training. ○ After the optimal 𝜆 ∗ is determined, the server generates a corresponding number of synthetic images (Alg. 1, Lines 6-9) and pretrains the LVM on the generated images dataset, then distributes LoRA-inserted model to devices (Alg. 1, Lines 10), as described in Section 3.3 and 3.4. 3 Federated Fine-Tuning. The main idea of FedFT is to de○ termine the LoRA configuration X𝑡 to maximize the performance gain, while balancing completion times across devices through a bandwidth allocation strategy B𝑡 . To handle device heterogeneity in FedFT, we design a contribution-aware LoRA configuration and bandwidth allocation (CLCBA) algorithm (Alg. 1, Lines 12-31). 𝑡 ) = ∇𝑓 (e 𝑡 Since the local gradient 𝑔𝑖 (e w𝑖,𝑙 𝑖 w𝑖,𝑙 ) captures the contribution of 𝑙-th layer to the loss function [18, 35], we use the L2 norm Í 𝑡 −1 )|| 2 of the last round 𝑡 − 1 to quantify the 𝐺𝑖𝑡 = 𝑙 ∈𝐿𝑡 ||𝑔𝑖 (e w𝑖,𝑙 2 𝑖 contribution of 𝑣𝑖 at round 𝑡. The performance gain is defined as the sum of all devices’ contributions divided by current round’s duration Í

𝑅𝑡 ≜

𝑡 𝑣𝑖 ∈V 𝐺𝑖 𝐻𝑡

Í

=

𝑣𝑖 ∈V

𝑡 −1 ) | | 2 e𝑖,𝑙 2 𝑙 ∈𝐿𝑖𝑡 | |𝑔𝑖 ( w 𝑡 𝐻

Í

.

(14)

We first initialize the depths of all devices as 1, i.e., set 𝑥𝑖𝑡 = 1 for all 𝑣𝑖 ∈ V (Line 15). In each iteration, we choose a device that obtains the maximum performance gain among all devices to add one LoRA layer. This process continues until the maximum performance gain is less than its value in the last iteration (Alg. 1, Line 17-28). Meanwhile, in each iteration of the process of Alg. 1, we call the water-filling based bandwidth allocation sub-algorithm (WFBA) to balance completion times across devices, which minimizes the completion time 𝐻 𝑡 in Eq. (14). The WFBA sub-algorithm is shown in Alg. 2. First, WFBA calcu𝑡,comp lates the maximum computation time 𝐻 max of all devices (Alg. 2,

Ma et al.

2 … 1

1 1

2 1 1 1 1

2 1 1 1

1 2 3 4 5

1 2 3 4 5

1 2 3 4 5

1 1

1

𝑅 =0.0026

𝑡

Computation time at depth 𝑥𝑖𝑡

𝐻𝑡

6 8

1 6 8 0 7 4

1 2 3 4 5

1 2 3 4 5

0 7 3

Communication time

Computation time after depth increase to 𝑥𝑖𝑡 + 1

Fig. 6: An example of CLCBA. Initially, devices 𝑣 1 −𝑣 5 have 𝑥𝑖𝑡 = 1. In the first iteration, increasing the depth of 𝑣 1 yields 𝑅𝑡 = 0.0026, and further increasing 𝑣 5 ’s depth raises 𝑅𝑡 to 0.0034. This continues until the maximum 𝑅𝑡 = 0.0102. Further increasing the depth of other devices will decrease 𝑅𝑡 to 0.0097, so CLCBA terminates and returns the LoRA configuration X𝑡 . In this process, WFBA is invoked to obtain B𝑡 , balancing the completion time of all devices. Lines 1-4). Since the waterline must exceed the maximum computa𝑡,comp 𝑡,comp tion time 𝐻 max , WFBA initializes the waterline as H0𝑡 = 2𝐻 max and iteratively solve its optimal value using Newton’s method [25] (Alg. 2, Lines 5-10). Finally, WFBA calculates the optimal bandwidth allocation 𝐵𝑖𝑡 for each device 𝑣𝑖 based on the converged waterline and returns the bandwidth configuration B𝑡 (Alg. 2, Lines 11-13). The intuitive diagram of CLCBA to obtain X𝑡 and B𝑡 are shown in Fig. 6.

6 Performance Evaluation 6.1 System Setup Environment Settings. We implement a FedFT experimental testbed comprising one server and 10 heterogeneous devices, as shown in Fig. 7. The server is equipped with an NVIDIA GeForce RTX 4090 GPU with 48GB GDDR6X. We use 1 Jetson AGX Xavier kit, 5 Jetson Orin NX kits, 3 Jetson AGX Orin kits, and 1 Jetson AGX Thor kit as heterogeneous devices. Particularly, we simulate more practical wireless communication in an edge environment instead of high-bandwidth laboratory networks, where communication time is typically negligible compared to computation time. Similar to [38], the server is at (0, 0), and each device is uniformly distributed within a circular area of 50 meters radius centered at coordinates (300, 0). The large-scale fading is represented as (𝑑/𝑑 0 ) −3.5 , where 𝑑 0 = 10 meters is the reference distance. The noise power is set as 𝜎 2 = 10−11 W, with a total system bandwidth of 𝐵 = 10MHz.

Devices

Server

 )H'L6\Q &$)) 6)/) )*/ *37)/

  







 

)H'L6\Q &$)) 6)/) )*/ *37)/

 









7LPHî2V



 )H'L6\Q &$)) 6)/) )*/ *37)/

  









7LPHî3V

Fig. 8: The test accuracy over different datasets (IID). Left: CIFAR-10; Middle: Caltech-101; Right: Food-101. [1] dataset. ViT-B/16 and ViT-L/16 consist of 86 and 307 million parameters, respectively. • CIFAR-10 [17] is composed of 60,000 color images, which are divided into 10 classes, each containing 6,000 images. The dataset is further split into 50,000 training images and 10,000 test images. • Caltech-101 [7] comprises 101 classes objects, totaling 9,144 samples. The number of samples per category ranges from 40 to 800 images, with each category averaging approximately 50 samples. We divide it into a 4:1 training-test split ratio. • Food-101 [1] contains a series of food datasets divided into 101 classes with high inter-class similarity. It comprises 101,000 samples, with each class including 750 training and 250 test samples. We analyze FedFT performance under both IID and non-IID data among devices. For the non-IID data case, we adopt the Dirichlet distribution [10] with 𝛼 dir = 0.3 to partition the dataset. In addition, we deploy the LLaMa-7B [36] to enrich prompts and StableDiffusion-v1-5 [31] to generate synthetic images as described in Section 3.3. Benchmarks. We compare FeDiSyn against two solutions with synthetic image pre-training (GPT-FL [45], FGL [43]) and two solutions deploying LoRA-based FedFT (SFLF [38], CAFF [29]). Specifically, GPT-FL and FGL generate images based on uploaded labels or text embeddings from devices, while SFLF and CAFF determine each device’s LoRA depth based on its capacity. For a fair comparison, FeDiSyn, SFLF, and CAFF adopt the same LoRA rank. • GPT-FL[45] is a synthetic image pre-training based approach that expands device-uploaded labels into prompts and generates images for pre-training LVM, while using FedAvg [24] for FedFT. • FGL[43] is a synthetic image pre-training based approach where devices use BLIP to generate prompts for local images, and the server employs these prompts to generate images for pre-training LVM, while utilizing FedAvg for FedFT. • SFLF[38] is a LoRA-based approach for FedFT, where the LVM is initialized with ImageNet parameters, and all devices fine-tune all transformer layers of the LVM. • CAFF[29] is an advanced LoRA-based approach that initializes LVM with ImageNet parameters, and assigns appropriate LoRA depths to each device to address system heterogeneity.

6.2

Fig. 7: Hardware devices of our testbed. Models and Datasets. We train ViT-B/16 [5] on CIFAR-10 [17] and Caltech-101 [7] datasets, and train ViT-L/16 [5] on Food-101



7LPHî3V

$FFXUDF\

𝑅 𝑡 =0.0015

𝑡

× 𝑅 =0.0097

$FFXUDF\

𝑅 𝑡 =0.0102 𝑅 𝑡 =0.0034

$FFXUDF\

ICPP ’26, September 28–October 1, 2026, Singapore

Evaluation Results

Training Performance. We evaluate the FedFT process of our FeDiSyn method and the benchmarks on the CIFAR-10, Caltech101 and Food-101 datasets under both IID and non-IID data distributions, as illustrated in Figs. 8 and 9. Overall, FeDiSyn consistently achieves the fastest convergence and highest accuracy on all datasets under both IID and non-IID data. For example, as shown in the third plot of Fig. 8, under IID data on Food-101, FeDiSyn reaches







7LPHî3V

 













  7LPHî2V

)H'L6\Q &$)) 6)/) )*/ *37)/

 









$FFXUDF\



  







$FFXUDF\





)H'L6\Q &$)) 6)/) )*/ *37)/

   







$FFXUDF\



 





$FFXUDF\





)H'L6\Q &$)) 6)/) )*/ *37)/

  

1HWZRUNWUDIILF0%

)H'L6\Q &$)) 6)/) )*/ *37)/





1HWZRUNWUDIILF0%

1HWZRUNWUDIILF0%

Fig. 10: The network traffic over different datasets (IID). Left: CIFAR-10; Middle: Caltech-101; Right: Food-101. 







$FFXUDF\



GPT-FL

FGL

SFLF

CAFF

FeDiSyn

Time Traffic

232.5 23243.5

220.3 21798.8

103.5 172.7

73.3 99.8

55.1 50.3

CIFAR-10 (Non-IID, 80%)

Time Traffic

348.6 31051.5

341.3 29840.5

203.7 254.0

161.1 146.0

92.3 93.5

Caltech-101 (IID, 70%)

Time Traffic

113.7 14547.6

96.4 12198.6

14.6 226.2

10.3 150.1

6.8 15.7

Caltech-101 (Non-IID, 65%)

Time Traffic

147.5 18564.2

116.7 14755.1

18.9 279.5

14.5 177.8

6.3 10.1

Food-101 (IID, 80%)

Time Traffic

453.8 30763.4

447.8 30375.9

418.9 457.3

319.1 290.1

175.8 42.5

Food-101 (Non-IID, 65%)

Time Traffic

284.9 13676.3

262.2 12002.8

423.1 399.5

328.4 262.1

124.7 7.4





)H'L6\Q &$)) 6)/) )*/ *37)/

  



7LPHî2V



)H'L6\Q &$)) 6)/) )*/ *37)/

 

 







  





$FFXUDF\

Fig. 11: The network traffic over different datasets (non-IID). Left: CIFAR-10; Middle: Caltech-101; Right: Food-101. 80% accuracy in 17,584s, while CAFF, SFLF, FGL, and GPT-FL require 31,905s, 41,885s, 44,783s, and 45,383s, respectively, meaning FeDiSyn reduces 44.9%, 58.0%, 60.7%, and 61.3% completion time compared to these benchmarks. Moreover, under non-IID data, FeDiSyn’s advantage becomes even more pronounced compared to other solutions. As shown in the third plot of Fig. 9, under nonIID data, FeDiSyn reduces 62.0%, 70.5%, 52.5%, and 56.2% time for reaching 65% accuracy compared to CAFF, SFLF, FGL, and GPT-FL, respectively. This is because FeDiSyn pre-training over synthetic images that enhances the generalization ability of initial model, alleviating the impact of non-IID data [12]. Network Traffic Consumption. In Fig. 10 and 11, we compare the network traffic consumption by different solutions to achieve certain accuracies on CIFAR-10 Caltech-101 and Food-101, respectively. By the results, FeDiSyn significantly reduces the network traffic compared with the benchmarks. For example, under non-IID data on Food-101, the consumed network traffic of FeDiSyn, CAFF, SFLF, FGL, and GPT-FL is 7.4MB, 262.1MB, 399.5MB, 12,002.8MB, and 13,676.3MB, respectively. FeDiSyn can reduce network traffic by 97.2%, 98.1%, 99.9%, and 99.9% compared with these methods. Compared with the LoRA-based methods CAFF and SFLF, FeDiSyn adopts a more efficient LoRA configuration strategy, thereby reducing network traffic when achieving the target accuracy. In addition, FGL and GPT-FL rely on full parameter FedFT, which incurs extremely high communication overhead. The complete experimental results of FedFT performance and network traffic consumption are provided in Table 2. Impact of Synthetic Image Volume. To estimate the values of coefficients 𝜅 1 − 𝜅 4 in Eq. (13), we fit this equation in a short calibration phase using a set of small values of 𝜆 via nonlinear least-squares (e.g., Levenberg-Marquardt method). Fig. 12 shows that Eq. (13) well describes the relationship among total training

 

1 = 2.452, 2 = 2.025, 3 = 0.191, 4 = 0.471





$FFXUDF\



 

   



 

1 = 8.780, 2 = 1.743, 3 = 0.159, 4 = 0.748







$FFXUDF\



 



1 = 2.083, 2 = 0.576, 3 = 0.445, 4 = 0.017

  



 







$FFXUDF\



Fig. 12: Fitting Eq. (13) with small 𝜆. Left: CIFAR-10; Middle: Caltech-101; Right: Food-101. *HQHUDWLRQ 3UHWUDLQLQJ )HG)7

7LPHî3V





1HWZRUNWUDIILF0%





1HWZRUNWUDIILF0%

1HWZRUNWUDIILF0%

)H'L6\Q &$)) 6)/) )*/ *37)/

Metric



7LPHî3V

Fig. 9: The test accuracy over different datasets (non-IID). Left: CIFAR-10; Middle: Caltech-101; Right: Food-101. 

Dataset CIFAR-10 (IID, 85%)

7LPHî2V



)H'L6\Q &$)) 6)/) )*/ *37)/

  











*HQHUDWLRQ 3UHWUDLQLQJ )HG)7

   











7LPHî3V





Table 2: The fine-tuning time (102 s) and network traffic (MB) to reach target accuracy.

7LPHî2V

)H'L6\Q &$)) 6)/) )*/ *37)/

ICPP ’26, September 28–October 1, 2026, Singapore

7LPHî2V







$FFXUDF\



$FFXUDF\

$FFXUDF\

Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models

*HQHUDWLRQ 3UHWUDLQLQJ )HG)7

   









6FDOLQJIDFWRU  6FDOLQJIDFWRU  6FDOLQJIDFWRU  Fig. 13: The time to reach target accuracy under different scaling factor 𝜆. Left: CIFAR-10 (85%); Middle: Caltech-101 (70%); Right: Food-101 (75%).

time, accuracy and the √︃ volume of generated data. Then we obtain 𝜅 +1

𝜅 3 𝐴𝑐𝑐

the optimal 𝜆 ∗ = 2 𝜅2𝑒ℎ0𝜅1 by the estimated coefficients. Fig. 13 shows the time to reach target accuracy under different scaling factor 𝜆 = 0.5𝜆 ∗, 𝜆 ∗, 1.5𝜆 ∗ and 2𝜆 ∗ on CIFAR-10, Caltech-101 and Food-101. 𝐴𝑐𝑐 is set as 85%, 70% and 80% for CIFAR-10, Caltech101 and Food-101, respectively. We observe that the generation and pre-training time increase linearly with 𝜆. For example, for Food101, the generation times are 5,050s, 10,100s, 15,150s, and 20,200s for 𝜆 = 0.5𝜆 ∗, 𝜆 ∗, 1.5𝜆 ∗ , and 2𝜆 ∗ , respectively. However, when the value of 𝜆 is too large or too small, the total completion time to achieve the target accuracy will be relatively long. For example, for Food-101, when 𝜆 increases from 0.5𝜆 ∗ to 2𝜆 ∗ , the time required to achieve 80% accuracy is 36,351.0s, 17,583.9s, 22,671.1s, and 27,943.9s, respectively. This is because insufficiently generated images weaken pre-training effectiveness for FedFT, whereas excessive images introduce substantial computational overhead, thereby prolonging image generation. The experimental results in Figs. 12-13 show that the values of 𝜆 ∗ obtained by the empirical model have the shortest completion time compared to other values. Ablation Study. FeDiSyn comprises two primary modules: 1) a feature-guided (FG) synthetic generation module with a diffusion model (DM), 2) a LoRA-based FedFT module deploying contributionaware LoRA configuration (CLC) and water-filling bandwidth allocation (WFBA). We conduct fine-grained ablation experiments

ICPP ’26, September 28–October 1, 2026, Singapore



)H'L6\Q )H'L6\QZR)* )H'L6\QZR'0 )H'L6\QZR&/& )H'L6\QZR:)%$

 





7LPHî3V

 )H'L6\Q )H'L6\QZR)* )H'L6\QZR'0 )H'L6\QZR&/& )H'L6\QZR:)%$

 







7LPHî2V

$FFXUDF\



$FFXUDF\

$FFXUDF\



Ma et al.

 )H'L6\Q )H'L6\QZR)* )H'L6\QZR'0 )H'L6\QZR&/& )H'L6\QZR:)%$



References













7LPHî3V



)H'L6\Q )H'L6\QZR)* )H'L6\QZR'0 )H'L6\QZR&/& )H'L6\QZR:)%$

 





$FFXUDF\



  

1HWZRUNWUDIILF0%



1HWZRUNWUDIILF0%

1HWZRUNWUDIILF0%

Fig. 14: Ablation study on accuracy. Left: CIFAR-10; Middle: Caltech-101; Right: Food-101. 

)H'L6\Q )H'L6\QZR)* )H'L6\QZR'0 )H'L6\QZR&/& )H'L6\QZR:)%$

  





$FFXUDF\





)H'L6\Q )H'L6\QZR)* )H'L6\QZR'0 )H'L6\QZR&/& )H'L6\QZR:)%$

   







$FFXUDF\



Fig. 15: Ablation study on Bandwidth. Left: CIFAR-10; Middle: Caltech-101; Right: Food-101. on CIFAR-10, Caltech-101, and Food-101 in Figs. 14 and 15 to evaluate the individual contribution of each sub-module. FeDiSyn consistently outperforms all variants, including those without FG, CLC, WFBA, or those utilizing GANs instead of DMs. For example, on CIFAR-10, FeDiSyn achieves 74% accuracy with 42.4%, 61.4%, 22.8%, and 41.7% time reduction and reduces network traffic usage by 47.0%, 65.6%, 12.5%, and 51.7%, compared to FeDiSyn w/o FG, FeDiSyn w/o DM, FeDiSyn w/o CLC, and FeDiSyn w/o WFBA, respectively. We also observe that the relative importance of synthetic pre-training grows with model scale. For example, for ViT-B/16 training on CIFAR-10 and Caltech-101, the performance gaps among different variants are relatively small, but for ViT-L/16 training on Food-101, FeDiSyn w/o CLC and w/o WFBA significantly outperform FeDiSyn w/o FG and w/o DM, even approaching the performance of FeDiSyn. This is because larger models possess a higher degree of representational complexity, making a high-quality pretraining for initialization far more critical than fine-tuning stage.

7

Grant BK20241453, and in part by the Fundamental Research Funds for the Central Universities under Grant 30925010408.

Conclusion

In this paper, we propose FeDiSyn, a unified framework that holistically considers both pre-training and federated fine-tuning (FedFT) of large vision models to address resource constraints, system heterogeneity, and non-IID data. FeDiSyn introduces a scaling law for synthetic pre-training under FedFT setting to determine the optimal number of synthetic images. A diffusion model is deployed on the server to generate synthetic images that align with the device’s local image label and feature distributions, thereby efficiently pretraining a superior initial global model for FedFT. During FedFT, we design a contribution-aware LoRA configuration and bandwidth allocation algorithm to enhance FedFT performance while addressing system heterogeneity. Experimental results indicate the superiority of our proposed framework compared with state-of-the-art solutions.

Acknowledgments The corresponding author of this paper is Junlong Zhou. This work was supported in part by the National Science Foundation of China (NSFC) under Grants 62402537, 62572244, 62402226 and U25A20442, in part by the Natural Science Foundation of Jiangsu Province under

[1] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101 – Mining Discriminative Components with Random Forests. In European conference on computer vision. 446–461. [2] Huancheng Chen and Haris Vikalo. 2024. Heterogeneity-guided client sampling: Towards fast and efficient non-iid federated learning. Advances in Neural Information Processing Systems 37 (2024), 65525–65561. [3] Hong-You Chen, Cheng-Hao Tu, Ziwei Li, Han Wei Shen, and Wei-Lun Chao. 2023. On the Importance and Applicability of Pre-Training for Federated Learning. International Conference on Learning Representations. [4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition. 248–255. [5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. International Conference on Learning Representations (ICLR) (2021). [6] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. 2024. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7382–7392. [7] Li Fei-Fei, Rob Fergus, and Pietro Perona. 2004. Learning Generative Visual Models from Few Training Examples: An Incremental Bayesian Approach Tested on 101 Object Categories. Computer Vision and Pattern Recognition Workshop (2004). [8] Tao Guo, Song Guo, and Junxiao Wang. 2024. Explore and Cure: Unveiling Sample Effectiveness With Context-Aware Federated Prompt Tuning. IEEE Transactions on Mobile Computing 23, 12 (2024), 14044–14054. [9] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. 2023. Is synthetic data from generative models ready for image recognition? International Conference on Learning Representations (ICLR) (2023). [10] Tzu-Ming Harry Hsu, Hang Qi, and Matthew Brown. 2019. Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification. arXiv preprint arXiv:1909.06335 (2019). [11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. International Conference on Learning Representations (ICLR) 1, 2 (2022), 3. [12] Divyansh Jhunjhunwala, Pranay Sharma, Zheng Xu, and Gauri Joshi. 2025. Initialization Matters: Unraveling the Impact of Pre-Training on Federated Learning. arXiv preprint arXiv:2502.08024 (2025). [13] Feiyang Kang, Newsha Ardalani, Michael Kuchnik, Youssef Emad, Mostafa Elhoushi, Shubhabrata Sengupta, Shang-Wen Li, Ramya Raghavendra, Ruoxi Jia, and Carole-Jean Wu. 2025. Demystifying Synthetic Data in LLM Pre-training: A Systematic Study of Scaling Laws, Benefits, and Pitfalls. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 10739–10758. [14] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models. arXiv:2001.08361 [15] Sai Praneeth Karimireddy, Martin Jaggi, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian U Stich, and Ananda Theertha Suresh. 2021. Breaking the centralized barrier for cross-device federated learning. Advances in Neural Information Processing Systems 34 (2021), 28663–28676. [16] Alex Krizhevsky et al. 2012. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012). [17] Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning multiple layers of features from tiny images. Citeseer. [18] Rukuo Li, Jianchun Liu, Hongli Xu, and Liusheng Huang. 2025. FedQuad: Adaptive Layer-wise LoRA Deployment and Activation Quantization for Federated FineTuning. arXiv preprint arXiv:2506.01001 (2025). [19] Jun Liu, Yunming Liao, Hongli Xu, Yang Xu, Jianchun Liu, and Chen Qian. 2025. Adaptive parameter-efficient federated fine-tuning on heterogeneous devices. IEEE Transactions on Mobile Computing (2025). [20] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. 2022. Pseudo Numerical Methods for Diffusion Models on Manifolds. In International Conference on Learning Representations. [21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV).

Scaling Synthetic-Image Pre-Training for Federated Fine-Tuning of Large Vision Models

[22] Qianpiao Ma, Yang Xu, Hongli Xu, Zhida Jiang, Liusheng Huang, and He Huang. 2021. FedSA: A Semi-Asynchronous Federated Learning Mechanism in Heterogeneous Edge Computing. IEEE Journal on Selected Areas in Communications 39, 12 (2021), 3654–3672. [23] Qianpiao Ma, Yang Xu, Hongli Xu, Jianchun Liu, and Liusheng Huang. 2024. FedUC: A Unified Clustering Approach for Hierarchical Federated Learning. IEEE Transactions on Mobile Computing 23, 10 (2024), 9737–9756. [24] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of Machine Learning Research, Vol. 54. 1273–1282. [25] Jorge J Moré and Danny C Sorensen. 1982. Newton’s method. Technical Report. Argonne National Lab.(ANL), Argonne, IL (United States). [26] John Nguyen, Kshitiz Malik, Maziar Sanjabi, and Michael Rabbat. 2023. Where to begin? exploring the impact of pre-training and initialization in federated learning. International Conference on Learning Representations (ICLR). [27] Tim Pearce, Tabish Rashid, Dave Bignell, Raluca Georgescu, Sam Devlin, and Katja Hofmann. 2024. Scaling Laws for Pre-training Agents and World Models. arXiv:2411.04434 [28] Danni Peng, Yuan Wang, Huazhu Fu, Jinpeng Jiang, Yong Liu, Rick Siow Mong Goh, and Qingsong Wei. 2025. Look Back for More: Harnessing Historical Sequential Updates for Personalized Federated Adapter Tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 19857–19865. [29] Kilian Pfeiffer, Mohamed Aboelenien Ahmed, Ramin Khalili, and Jörg Henkel. 2024. Efficient Federated Finetuning of Tiny Transformers with ResourceConstrained Devices. arXiv preprint arXiv:2411.07826 (2024). [30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning (PMLR), Vol. 139. 8748–8763. [31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695. [32] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. 2015. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning. pmlr, 2256–2265. [33] Shangchao Su, Bin Li, and Xiangyang Xue. 2024. Fedra: A random allocation strategy for federated tuning to unleash the power of heterogeneous clients. In European Conference on Computer Vision. Springer, 342–358. [34] Guangyu Sun, Matias Mendieta, Jun Luo, Shandong Wu, and Chen Chen. 2023. FedPerfix: Towards Partial Model Personalization of Vision Transformers in Federated Learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 4988–4998. [35] Yuchang Sun, Yuexiang Xie, Bolin Ding, Yaliang Li, and Jun Zhang. 2025. Exploring Selective Layer Fine-Tuning in Federated Learning. In IEEE International Symposium on Information Theory (ISIT). [36] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023). [37] Shiqiang Wang, Tiffany Tuor, Theodoros Salonidis, Kin K Leung, Christian Makaya, Ting He, and Kevin Chan. 2019. Adaptive Federated Learning in Resource Constrained Edge Computing Systems. IEEE Journal on Selected Areas in Communications 37, 6 (2019), 1205–1221. [38] Zixin Wang, Yong Zhou, Yuanming Shi, and Khaled B. Letaief. 2025. Federated Fine-Tuning for Pre-Trained Foundation Models Over Wireless Networks. IEEE Transactions on Wireless Communications 24, 4 (2025), 3450–3464. [39] Chulin Xie, De-An Huang, Wenda Chu, Daguang Xu, Chaowei Xiao, Bo Li, and Anima Anandkumar. 2024. Perada: Parameter-Efficient Federated Learning Personalization with Generalization Guarantees. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 23838–23848. [40] Yunlu Yan and Lei Zhu. 2023. A simple data augmentation for feature distribution skewed federated learning. arXiv preprint arXiv:2306.09363 (2023). [41] Mingzhao Yang, Shangchao Su, Bin Li, and Xiangyang Xue. 2024. Exploring oneshot semi-supervised federated learning with pre-trained diffusion models. In Proceedings of the AAAI conference on artificial intelligence, Vol. 38. 16325–16333. [42] Biao Zhang, Zhongtao Liu, Colin Cherry, and Orhan Firat. 2024. When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method. In The Twelfth International Conference on Learning Representations. [43] Jie Zhang, Xiaohua Qi, and Bo Zhao. 2023. Federated generative learning with foundation models. arXiv preprint arXiv:2306.16064 (2023). [44] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. Adaptive Budget Allocation for ParameterEfficient Fine-Tuning. In The Eleventh International Conference on Learning Representations (ICLR).

ICPP ’26, September 28–October 1, 2026, Singapore

[45] Tuo Zhang, Tiantian Feng, Samiul Alam, Dimitrios Dimitriadis, Sunwoo Lee, Mi Zhang, Shrikanth S Narayanan, and Salman Avestimehr. 2025. Gpt-fl: Generative pre-trained model-assisted federated learning. In Proceedings of the Computer Vision and Pattern Recognition Conference. 1761–1770. [46] Zhuo Zhang, Yuanhang Yang, Yong Dai, Qifan Wang, Yue Yu, Lizhen Qu, and Zenglin Xu. 2023. FedPETuning: When federated learning meets the parameterefficient tuning methods of pre-trained language models. In Association for Computational Linguistics. 9963–9977. [47] Tianfei Zhou, Ye Yuan, Binglu Wang, and Ender Konukoglu. 2024. Federated Feature Augmentation and Alignment. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 12 (2024), 11119–11135. [48] Guangxu Zhu, Dongzhu Liu, Yuqing Du, Changsheng You, Jun Zhang, and Kaibin Huang. 2018. Towards an intelligent edge: Wireless communication meets machine learning. arXiv preprint arXiv:1809.00343 (2018).

Related documents

• CLEAR-MoE: Shared-Basis Expert Extraction from Frozen Vision Transformers via Calibration-Driven Layer Selection#321816
• Pretraining Reusable Inference Across Views with Synthetic Task Priors#467117
• Effective Synthetic Image Detection via Noise Residual Clustering#363183
• Natural Image Autoencoder-Based fMRI Representations for Trait and State Prediction#1119240
• Learning from Distributed Eyes: Leveraging Collaborative Perception for Automated Model Adaptation#1012693