💻 computer science

Incentivizing Multi-Tenant Split Federated Learning for Foundation Models at the Network Edge

This paper proposes PRINCE, a novel price-incentive mechanism that coordinates self-interested devices across multiple Split Federated Learning tenants to efficiently fine-tune diverse foundation models at the network edge by combining bias-resilient aggregation, convergence-guided contribution evaluation, and Stackelberg equilibrium-based strategy optimization.

Original authors: Songyuan Li, Jia Hu, Geyong Min, Haojun Huang

Published 2026-01-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Songyuan Li, Jia Hu, Geyong Min, Haojun Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where giant, super-smart AI brains (called Foundation Models) exist, like a super-intelligent chef who knows how to cook every dish in the world. But this chef is too big and expensive to fit into a single kitchen (your phone or smartwatch). To teach this chef a specific new recipe (like "how to make the perfect pizza" or "how to understand Turkish accents"), you need to train it.

Usually, training these giants requires massive computers that regular people don't have. Also, you don't want to send your private photos or voice notes to a central server to do the training; you want to keep that data on your own device.

The Problem: The "Tug-of-War" for Helpers
This paper introduces a system called Split Federated Learning (SFL). Think of it as a team effort:

You (the local device) do the easy, preliminary prep work (like chopping vegetables).
The Edge Server (a powerful computer nearby) does the heavy lifting (like cooking the complex sauce).
You never send your raw ingredients (private data) to the server; you only send the chopped veggies (intermediate results).

However, there's a catch. In the real world, there isn't just one chef needing help. There are multiple tenants (different companies or apps) all trying to train their own specific AI models at the same time using the same pool of local devices (your phone, your watch, your smart speaker).

Tenant A wants to train an AI to recognize food images.
Tenant B wants to train an AI to analyze text sentiment.
Tenant C wants to train an AI to transcribe speech.

All these tenants are competing for your device's attention. But your device is "selfish" (in a smart way): it doesn't want to work for free. It has a battery and limited power. If Tenant A offers \ $1 and Tenant B offers \$ 5, your device will naturally choose Tenant B.

The paper argues that existing methods are bad at this. They either ignore the competition between tenants or fail to pay the right devices enough to get the best results. This leads to two problems:

Bias: If only a few specific devices do all the work, the AI model becomes biased (like a chef who only learned to cook from one specific region's recipes).
Inefficiency: The AI takes too long to learn because the wrong devices are working, or not enough of them are working.

The Solution: PRINCE (The Price-Incentive Mechanism)
The authors propose a new system called PRINCE (Price-Incentive Mechanism). Here is how it works, using a simple analogy:

1. The Auction House (The Game)
Imagine a marketplace where multiple landlords (the Tenants) are trying to rent out their apartments to a group of tenants (the Devices).

The Landlords want the best tenants to keep their buildings clean and happy.
The Tenants want the highest rent but don't want to do too much work.
PRINCE acts as the smart auctioneer. It calculates exactly how much each Landlord should pay each Tenant to get the perfect balance. It ensures that no Landlord is left out and that the Tenants are motivated to work hard.

2. The "Fairness" Rule (Bias-Resilient Aggregation)
In the past, if a device only participated 50% of the time, its contribution was often ignored or counted incorrectly, messing up the final AI model.

PRINCE's Trick: It uses a special math formula (like a "fairness filter") that says, "Even if you only showed up half the time, we will count your contribution as if you were there the whole time, but we'll adjust the math so the final result isn't skewed." This ensures the AI learns correctly even if devices are busy or only join occasionally.

3. The "Contribution" Score
Before the training even starts, PRINCE predicts how much a specific device will help. It doesn't wait until the end to see who was good; it estimates it upfront based on the device's power and data quality. This helps the Landlords (Tenants) decide exactly how much to pay to get the best helpers.

The Results: Speed and Fairness
The authors tested this with four different types of AI tasks (recognizing food images, analyzing text, transcribing speech, and answering questions) using 100 simulated devices.

Speed: PRINCE made the AI training 3 times faster than the best existing methods. It's like going from a slow, winding country road to a high-speed highway.
Accuracy: It consistently hit the target goals (like 90% accuracy in image recognition) where other methods failed or took forever.
Balance: It ensured that all four different "Landlords" got a fair share of the "Tenants," so no single AI model was left behind while another got all the help.

In a Nutshell
This paper solves the problem of "too many cooks in the kitchen." It creates a smart payment system that lets multiple AI projects compete fairly for the help of your personal devices. By paying the right amount to the right devices and using a special math trick to keep the learning unbiased, it makes training giant AI models on the edge of the network (your local devices) much faster, fairer, and more efficient.

Technical Summary: Incentivizing Multi-Tenant Split Federated Learning for Foundation Models at the Network Edge

Problem Statement
Foundation Models (FMs) like GPT-4, BERT, and LLaMA require extensive fine-tuning on domain-specific data to achieve high performance in downstream tasks. However, local devices (e.g., smartphones, smartwatches) often lack the computational resources to handle full-parameter FM fine-tuning, while privacy concerns prevent data from leaving the device. Split Federated Learning (SFL) addresses this by offloading partial computations to edge servers, enabling device-edge synergistic fine-tuning.

While existing SFL research focuses on single-tenant scenarios, practical edge networks increasingly host multiple SFL tenants competing for the same pool of local devices to support diverse downstream tasks (e.g., image classification, sentiment analysis, speech-to-text). This multi-tenant environment introduces three critical challenges:

Independent Device Participation: Self-interested local devices select tasks based on incentives, potentially causing data bias where specific tasks are dominated by a subset of devices.
Device Contribution Assessment: It is difficult to predict how a device's participation level impacts the final FM performance before the training process is complete, hindering the design of effective incentive strategies.
Inter-Tenant Competition: Multiple tenants compete for high-quality devices, requiring a mechanism to balance participation and ensure fair performance across all tenants.

Current incentive mechanisms are largely designed for single-tenant Federated Learning (FL) or standard SFL and lack the tailored strategies needed to coordinate self-interested devices across competing multi-tenant FM fine-tuning tasks.

Methodology
The authors propose PRINCE (Price-Incentive Mechanism), a novel framework designed to guide multiple SFL tenants in offering strategic price incentives to solicit high-quality device participation. The methodology is built upon three core components:

Bias-Resilient Global SFL Aggregation:
To address independent device participation, the paper introduces an unbiased global model aggregation scheme. Unlike standard FedAvg which assumes full participation, this scheme re-weights model updates from participating devices by the inverse of their participation probabilities ( $1/q_{i,j}$ ). This ensures that the global FM model remains unbiased toward the scenario of full device participation, even when devices join tasks probabilistically.
SFL Convergence Bound for Contribution Assessment:
To solve the challenge of assessing device contributions without completing the training, the authors derive a rigorous SFL convergence bound. This bound analytically relates the expected global FM loss to device participation levels ( $q_{i,j}$ ), local data heterogeneity, and gradient magnitudes. This theoretical bound serves as a surrogate for the expected global loss, allowing SFL tenants to estimate the impact of different pricing strategies on FM performance prior to execution.
Stackelberg and Congestion Game Modeling:
The interaction between SFL tenants and local devices is modeled as a Multi-Leader Multi-Follower Stackelberg Game:
- Leaders (SFL Tenants): Determine optimal pricing strategies ( $P_i$ ) to minimize their expected global FM loss within a payment budget.
- Followers (Local Devices): Respond by selecting participation levels ( $q_{i,j}$ ) for various tasks to maximize their utility (payment minus computation cost).
- Inter-Tenant Competition: The competition among tenants for shared devices is modeled as a Congestion Game. Tenants iteratively adjust their pricing strategies to improve their global loss, converging toward a Nash Equilibrium (NE) within the congestion game, which corresponds to the Stackelberg Equilibrium (SE) of the overall system.

A decentralized algorithm is proposed to implement this game, where tenants and devices iteratively update strategies until the SE is reached.

Key Contributions

First Exploration of Multi-Tenant SFL for FMs: The paper is the first to explore effective incentive mechanisms for multi-tenant SFL specifically tailored for Foundation Model fine-tuning across diverse data modalities (text, images, audio).
Bias-Resilient Design: It introduces a global aggregation scheme that eliminates model biases caused by independent, probabilistic device participation.
Theoretical Convergence Bound: The derivation of a rigorous SFL convergence bound provides an analytical tool for tenants to evaluate device contributions and design pricing strategies without prior knowledge of the final model performance.
PRINCE Mechanism: The development of a decentralized price-incentive algorithm based on Stackelberg and congestion game theory that achieves finite-time convergence to an optimal equilibrium, balancing inter-tenant competition and device incentives.

Experimental Results
The authors evaluated PRINCE using simulations with four representative SFL tenants (ViT for image classification, BERT for sentiment analysis, Whisper for speech-to-text, and LLaMA for question answering) across 100 local devices.

Performance Acceleration: PRINCE accelerated FM fine-tuning by up to 3.07x compared to state-of-the-art approaches (FedPEFT, FAIR, and MSDA).
Target Accuracy: PRINCE consistently met performance targets for all downstream tasks. In contrast, benchmarks like FAIR and MSDA failed to reach target accuracies for specific tasks (e.g., Question Answering and Speech-to-Text) within reasonable deadlines.
Comparison with FL: Compared to classical FL (FedPEFT), PRINCE achieved target accuracy 31.36% faster on average, highlighting the efficiency of the device-edge synergy in SFL for resource-constrained FM workloads.
Robustness: The mechanism maintained balanced optimization across tenants and demonstrated finite-time convergence (averaging 42 iterations) regardless of system scale.

Significance
The paper claims that PRINCE provides a solid foundation for the widespread adoption of SFL in edge networks for Foundation Model fine-tuning. By addressing the specific complexities of multi-tenant environments—specifically independent device behavior, contribution assessment, and inter-tenant competition—the proposed mechanism ensures that diverse downstream tasks can be efficiently and fairly executed while preserving data privacy. The work bridges the gap between theoretical incentive design and the practical requirements of deploying large-scale, privacy-preserving AI at the network edge.

More like this