Research map / Federated learning and privacy / Personalized and heterogeneous FL
Multi-Session Budget Optimization for Forward Auction-based Federated Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MUSBO, a hierarchical RL (two DQN agents) budget optimization strategy for model users bidding in multi-session forward auction-based federated learning, evaluated in simulated auctions on six image datasets.
Problem
Model users in forward auction-based federated learning must recruit data owners over multiple training sessions under a single budget, but existing AFL bidding strategies only optimize within one session and cannot perform budget pacing across sessions.
Approach
Hierarchical RL with an inter-session budget pacing agent that sets each session's budget from historical outcomes and an intra-session bidding agent that sets per-DO bid prices within that budget, both trained with DQN to maximize total reputation-based utility.
Key results
- MUSBO achieves on average 12.28% higher utility and 14.52% more data acquired than the best baseline across budgets and datasets in the IID setting.
- Test accuracy improves by 1.23% on average over the best baseline in the IID no-noise setting.
- In the Non-IID with noise scenario, MUSBO gains 1.49% (10% noisy) and 1.72% (25% noisy) accuracy over the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-session budget pacing and intra-session bidding instead of the best baseline | Model users in multi-session forward auction-based federated learning · Six benchmark datasets, IID data with varying dataset sizes | utility: 12.28% higher utility | Helped |
| Use multi-session budget pacing and intra-session bidding instead of the best baseline | Model users in multi-session forward auction-based federated learning · Six benchmark datasets, IID data with varying dataset sizes | #data: 14.52% improvement | Helped |
| Use multi-session budget pacing and intra-session bidding instead of the best baseline | Federated models trained from auction-recruited data owners · Six benchmark datasets, IID data with varying dataset sizes | test accuracy: 1.23% on average | Helped |
| Use multi-session budget pacing and intra-session bidding instead of the best baseline | Model users in multi-session forward auction-based federated learning · IID data, same-size DO datasets with noisy samples | utility: 2.41% | Helped |
| Use multi-session budget pacing and intra-session bidding instead of the best baseline | Federated models trained from auction-recruited data owners · IID data, same-size DO datasets with noisy samples | accuracy: 1.27% | Helped |
| Use multi-session budget pacing and intra-session bidding instead of the best baseline | Federated models trained from auction-recruited data owners · Non-IID data with 10% noisy data | accuracy: 1.49% higher | Helped |
| Use multi-session budget pacing and intra-session bidding instead of the best baseline | Federated models trained from auction-recruited data owners · Non-IID data with 25% noisy data | accuracy: 1.72% higher | Helped |
Surprising findings
- RLB and the DQN-based method outperform the static bidding functions (BM, FBs, FBc), which the authors attribute to adaptivity in dynamic auction environments.
Limitations
- Evaluation relies on simulated AFL auctions with no public AFL bidding dataset, so realism is unverified.
- Only nine simulated MUs compete, and experiments use image classification with small MLP/VGG11 models.
Experimental setup
Datasets: MNIST CIFAR-10 Fashion-MNIST EMNIST-digits EMNIST-letters KMNIST
Benchmarks: MNIST CIFAR-10 Fashion-MNIST EMNIST-digits EMNIST-letters KMNIST
Metrics: number of data samples won (#data) utility (cumulative reputation) test accuracy of resulting FL model
Baselines: Constant bid Randomly generated bid Below max utility bid Linear-form bid Bidding Machine Fed-Bidder (FBs and FBc) RLB
Builds on: DQN Fed-Bidder FedAvg Beta Reputation System Shapley value
Closest papers by meaning
- Reputation-aware Revenue Allocation for Auction-based Federated Learning (AAAI 2025 · similarity 0.91)
- DualGFL: Federated Learning with a Dual-Level Coalition-Auction Game (AAAI 2025 · similarity 0.87)
- Pursuing Overall Welfare in Federated Learning through Sequential Decision Making (ICML 2024 · similarity 0.84)
- FedRSPO+: A Heterogeneity-aware Algorithm for Decision-focused Federated Learning (NeurIPS 2026 · similarity 0.84)
- FedCross: Intertemporal Federated Learning Under Evolutionary Games (AAAI 2025 · similarity 0.84)
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.84)
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.