Source-linked AI summary
Toward an Automated Auction Framework for Wireless Federated Learning Services Market
Yutao Jiao, Ping Wang, Dusit Niyato, Bin Lin, Dong In Kim
TL;DR
Centralized training raises privacy concerns, while federated learning needs enough data owners and resources to train models effectively. The paper proposes two auction mechanisms for a wireless federated learning market, combining an approximate RMA design with an automated DRLA mechanism based on deep reinforcement learning and graph neural networks. The mechanisms target social welfare while incorporating data quality and wireless channel conflicts, and are designed to be truthful and individually rational.
Problem
The paper addresses how to organize wireless federated learning participation and resource allocation while preserving privacy, handling data heterogeneity and channel conflicts, and maximizing social welfare.
Method
The paper designs an approximate reverse multi-dimensional auction and an automated auction using deep reinforcement learning and graph neural networks.
Results
The proposed auction mechanisms efficiently maximize social welfare in simulations, with DRLA showing significant improvement compared with RMA.
Takeaways & Limitations
The mechanisms provide practical trading strategies for coordinating data owners’ data and computing resources in federated training.
Abstract
from arXiv · showhide
In traditional machine learning, the central server first collects the data owners' private data together and then trains the model. However, people's concerns about data privacy protection are dramatically increasing. The emerging paradigm of federated learning efficiently builds machine learning models while allowing the private data to be kept at local devices. The success of federated learning requires sufficient data owners to jointly utilize their data, computing and communication resources for model training. In this paper, we propose an auction based market model for incentivizing data owners to participate in federated learning. We design two auction mechanisms for the federated learning platform to maximize the social welfare of the federated learning services market. Specifically, we first design an approximate strategy-proof mechanism which guarantees the truthfulness, individual rationality, and computational efficiency. To improve the social welfare, we develop an automated strategy-proof mechanism based on deep reinforcement learning and graph neural networks. The communication traffic congestion and the unique characteristics of federated learning are particularly considered in the proposed model. Extensive experimental results demonstrate that our proposed auction mechanisms can efficiently maximize the social welfare and provide effective insights and strategies for the platform to organize the federated training.
I. INTRODUCTION
The paper motivates a wireless federated learning services market to address privacy concerns and recruit sufficient data owners. It proposes two truthful auction mechanisms that account for federated learning characteristics and aim to maximize social welfare.
- Motivation: Federated learning keeps private data on local devices while collaboratively training a shared model, addressing concerns about centralized data collection.
- Motivation: Successful federated learning requires broad participation from data owners contributing sufficient training data.
- Market design: The platform formulates worker selection, spectrum allocation, and payment design as a social welfare maximization problem under wireless constraints.
- Auction mechanisms: The proposed RMA mechanism approximates social welfare maximization using data owners’ bids and resource attributes.
- Auction mechanisms: The DRLA mechanism combines deep reinforcement learning and graph neural networks to improve social welfare while representing wireless channel conflicts with a conflict graph.
- Auction mechanisms: Both auction mechanisms are theoretically proved strategy-proof, meaning truthful and individually rational.
A. Preliminary Knowledge of Federated Learning
Federated learning trains local models on distributed devices and repeatedly aggregates their updates into a global model. The described process uses local optimization followed by platform-side federated averaging until convergence.
- Federated learning process: The platform initializes the task, model hyper-parameters, training process, and initial global model before communicating them to workers.
- Federated learning process: Workers update local model parameters using their private datasets and devices during each global training epoch.
- Federated learning process: The platform receives and aggregates workers’ local models, then sends updated global parameters for the next training round.
- Federated learning process: Training repeats until the global loss converges and can support methods including SVMs, convolutional neural networks, and linear regression.
- Federated averaging: FedAvg is the aggregation algorithm used to combine local models into the global model.
B. Local Data Evaluation
The platform evaluates local data using data size and distribution, with EMD capturing divergence from a reference distribution and a fitted utility function linking these factors to model quality.
- Local data evaluation accounts for collection, sensing, annotation, cleaning, redundancy elimination, and anomaly-detection costs.
- Data quality is characterized by local data size and distribution because local FL data are typically non-IID.The platform uses data size and EMD to quantify potential contributions to task completion.
- EMD quantifies weights divergence, with larger values adversely affecting global model quality and prediction accuracy.
- The actual distribution Pa can serve as the reference distribution when supplied by public knowledge or estimated from the platform’s historical data.
- The fitted model-quality relationship uses workers’ total data size D and average EMD ∆ to represent model quality q.The platform’s data utility ϕ is then defined as a linear function of q.
- The utility function models performance degradation from increasing average EMD and diminishing marginal returns from increasing total data size.The fitting parameters κ1 through κ6 are positive, and σmax bounds the acceptable EMD range for service quality.
C. Auction based FL Services Market
The wireless FL services market recruits workers through an auction that accounts for private service costs, data quality, computation, communication, and channel allocation constraints.
- The platform conducts an auction to recruit enough qualified workers for successful federated training.The auction is illustrated as supporting the FL process.
- The platform announces the task and auction rule, then data owners report bids, resource information, and requested wireless channels.A bid reveals the private service cost or valuation, while the type profile also includes data and channel-related information.
- The model excludes adverse attacks and assumes data owners cannot claim higher data quality than they truly possess.Reporting higher data size or lower EMD would constitute model update poisoning under the stated assumption.
- The platform selects workers and notifies all data owners of the worker set and corresponding payments.Workers are single-minded in channel allocation and accept their requested channel set only if they win.
- Data owners estimate service costs from local data, computation, and communication expenses.The local data cost includes sensing, annotation, cleaning, redundancy elimination, and anomaly detection; computation cost is linear in data size.
- The communication model uses Shannon’s formula with FDMA channels, bandwidth, channel gains, transmission time, and unit energy costs.The model assumes slow-fading stable channels, ignores communication overhead, and assumes channel conditions are known to both parties.
- The reported bid equals the true service cost because the proposed auction mechanisms are truthful.
E. Social Welfare Optimization and Desired Economic Properties
The paper formulates social welfare as the combined utility of the platform and data owners, subject to nonconflicting channel allocation, and requires truthful, individually rational, polynomial-time mechanisms.
- The platform’s utility equals data utility minus total platform cost and payments to workers.
- Each data owner’s utility is its payment minus its service cost.
- Social welfare is defined as the sum of the platform’s utility and the data owners’ utilities, and the mechanism maximizes this quantity.
- Allocated channel sets must be disjoint across selected workers under the FDD communication scheme.
- Truthfulness requires that no data owner can gain utility by reporting a false type while other owners’ types remain fixed.
- Individual rationality requires every data owner’s utility to be nonnegative, while computational efficiency requires polynomial-time auction completion.
IV. REVERSE MULTI-DIMENSIONAL AUCTION MECHANISM FOR FEDERATED TRAINING
The RMA is a randomized, greedy reverse multi-dimensional auction that partitions data owners by EMD, selects compatible workers, and determines critical payments. It is designed to maximize social welfare while providing truthfulness, individual rationality, and polynomial-time computation.
- Mechanism design: The RMA divides data owners into EMD-based groups, then performs worker selection and service payment determination in three consecutive phases.Each group covers an EMD interval, replacing reported EMD values with interval midpoints for the sub-auctions.
- Worker selection: Within each group, the mechanism excludes channel-conflicting owners and greedily sorts compatible owners by non-increasing marginal social welfare density.Previously selected workers from other groups are also used when determining conflicts and marginal densities.
- Payments: The RMA sets a winning worker’s payment using a critical bid determined by the marginal-density threshold at which the worker would cease to win.The payment calculation re-executes worker selection after removing the worker to identify the relevant critical owner or threshold.
- Theoretical properties: The RMA is universally truthful and individually rational, with winning workers receiving payments at least equal to their costs.Truthfulness follows from monotone selection and critical payments, while individual rationality follows from non-negative worker utility.
- Computational efficiency: The mechanism runs in polynomial time O(N 3), because the total sub-auction complexity is bounded by the sum of O(|Θj|3) across groups.The cubic bound is obtained from the sub-auction loop complexity and the group-size inequality.
V. DEEP REINFORCEMENT LEARNING BASED AUCTION MECHANISM (DRLA)
The RMA’s social welfare is restricted by randomization and inadequate representation of channel conflicts. The DRLA mechanism uses AI to improve social welfare while preserving incentive compatibility and individual rationality.
- Motivation: The DRLA mechanism is motivated by limitations of RMA randomization and insufficient exploitation of channel conflicts in worker selection.The proposed approach seeks higher social welfare without sacrificing IC and IR.
A. Feature engineering with embeddings of wireless spectrum conflict graph
The feature-engineering stage represents wireless channel conflicts with a graph and converts discrete channel information into continuous node and graph embeddings using a GCN. These embeddings are combined with owner types, channel data, and auction states.
- Conflict graph: A conflict graph models each data owner as a node and each undirected edge as a wireless-channel conflict between connected owners.Owners requesting overlapping channels, such as channel 6, are connected by a conflict edge.
- Graph embeddings: A multi-layer GCN maps discrete channel information and graph structure into continuous embeddings for individual data owners.The final-layer output provides one node embedding vi for each owner.
- Graph embeddings: Each owner’s final embedding concatenates its node embedding with a graph embedding representing the overall conflict graph.The resulting representation is used alongside the owner’s original features.
- Auction features: The DRL input feature combines bids, data size, EMD, requested channels, channel information, learned embeddings, and the current worker-selection state.The binary state indicates whether the owner has already been selected as a worker.
B. Automated mechanism under deep Q-learning framework
The automated mechanism uses deep Q-learning to select non-conflicting workers greedily according to learned action values. Its state, reward, policy, and training design align learned evaluation with social-welfare maximization while preserving monotonicity properties.
- Deep Q-learning framework: At each step, deep Q-learning selects a non-conflicting data owner that maximizes the evaluation function Q.Selection continues greedily until termination conditions are reached.
- Training: DRLA training samples owner features from a known distribution or real dataset and updates the evaluation network with experience replay, SGD, and a periodically reset target network.Double deep Q-learning is used to address overestimation from selecting and evaluating actions with the same Q function.
- States and actions: The state records which owners are already in the candidate worker set, while each action selects an unchosen owner node.The state transition adds the selected owner to the candidate set.
- Rewards: Each reward equals the increase in social welfare from an action, so cumulative reward equals the final achieved social welfare.The reward design directly aligns the learning objective with the auction’s optimization target.
- Policy network: The evaluation network estimates action quality from state-action pairs and uses a monotonic architecture that decreases with bids and EMD while increasing with data size.The architecture uses positive coefficients to enforce these monotonic relationships.
- Auction properties: The DRLA payment is the worker’s critical bid, and the trained greedy policy provides monotonic worker selection that supports truthfulness and individual rationality.The critical bid is computed by comparing the worker’s score with the score of a critical data owner.
VI. EXPERIMENTAL RESULTS AND DISCUSSIONS
The experiments evaluate the proposed data utility function and the RMA and DRLA auction mechanisms, examining their performance through federated-learning simulations.
- The experiments first verify the proposed data utility function and then evaluate the RMA and DRLA mechanisms.
A. Verification for Data Utility Function
A CNN-based MNIST federated-learning experiment with two FedAvg workers tests whether the proposed data utility function matches observed prediction accuracy. The fitted function is reported to agree well with the real experimental results.
- A CNN on MNIST uses two FedAvg workers to test the data utility function through federated handwritten-digit recognition.
- 100-instance averages vary total data size and average EMD by changing workers’ local data sizes and label counts.
- The data utility function fits the real experiment results well, with prediction accuracy used as the data utility.
B. Performance of RMA and DRLA mechanisms
Simulations compare RMA and DRLA while varying auction and federated-learning parameters, including participants, data requirements, EMD limits, and RMA group count. Results show welfare gains alongside fairness trade-offs.
- The simulations compare the manually designed RMA and automated DRLA strategyproof auction mechanisms.
- DRLA convergence is evaluated using held-out validation performance.
- Social welfare and selected workers increase with the number of data owners at a diminishing rate for both mechanisms.
- DRLA achieves higher social welfare than RMA but is less fair because it more effectively selects owners with larger marginal social welfare.
- The benchmark spectrum auction achieves lower social welfare than the proposed mechanisms but provides better fairness.
- Raising the required data size and lowering the allowed EMD increase social welfare when enough data owners satisfy the requirements.
VII. CONCLUSION
The paper proposes wireless federated-learning auctions that value data quality and wireless resources, combining RMA with DRLA to improve social welfare while retaining truthfulness and individual rationality.
- RMA accounts for bids, data size, EMD, and wireless channel demand when allocating federated-learning services.
- DRLA uses deep reinforcement learning and a graph neural network to determine service allocation and payment automatically.
- Both RMA and DRLA possess truthfulness and individual rationality, while the data quality function is verified through real-world experiments.