Robust Bidding Strategies under Censored Feedback for Auction-Based Federated Learning

Auction-based Federated Learning (AFL) provides a principled framework for incentivizing self-interested data owners (DOs) to participate in collaborative learning initiated by data consumers (DCs) through market mechanisms. A central challenge in AFL is to determine how a budget-constrained DC should bid for data-use rights. Existing methods often conflate the submitted bid with the realized payment. This treatment is inconsistent with the widely adopted second-price sealed-bid (SPSB) mechanism, in which the highest bidder wins but pays the market price determined by the competing bids rather than its own bid. Moreover, the market price is observed by a target DC only after a win; after a loss, the DC observes only that the market price exceeds its submitted bid, producing right-censored feedback. To address these challenges, we propose KMM-AFL, a return-on-investment (ROI)-aware bidding framework that explicitly separates bid price from expected SPSB payment and combines DO utility estimation with request-conditioned market-price modeling. KMM-AFL uses a conditional Kaplan Meier (KM) construction to exploit both winning and losing auction records. Because exact repetitions of high-dimensional bid requests are sparse, a conditional Markov Network estimates the observed winning-price and losing-probability terms required by KM. The recovered request-conditioned market-price distribution and the estimated DO valuation are then used to select the largest bid satisfying a common expected-return requirement and the remaining-budget constraint. Extensive evaluations across 6 benchmark datasets show that KMM-AFL outperforms state-of the-art baselines, achieving a 16.56% increase in DC utility and a 3.36% improvement in FL model accuracy.