Privacy-Preserving Federated Learning for Corruption Risk Detection in Public Procurement

Abstract

The detection of elevated corruption risk in public procurement, where risk is measured by objective procedural indicators rather than by confirmed corruption, is typically constrained by the fragmentation of the underlying records: each national authority holds its own, and legal obligations prevent them from being pooled. The current study examines whether authorities can improve detection by collaborating without exchanging data records. To this end, a federated framework is developed in which each authority trains locally and publishes only model parameters and anonymized class-conditional statistics to a coordination layer. It is evaluated on real procurement records from three European Union member states, using a predictor set audited to exclude attributes that determine the label. Collaboration improved detection substantially over independent operation across the federation, recovering most of the performance achievable by centralized training, with the largest improvements at the operating points relevant to bounded audit capacity. A model trained on two authorities and applied to a third performed near the local base rate, so participation is a condition of benefit rather than an optional route to it. Perturbing the exchanged quantities carries a measurable cost that grows with model size, so that beyond a modest noise level, an interpretable linear model outperforms a higher-capacity network.