BRAID neural network predicts security game equilibria 43x faster

Computing Nash equilibria in interdependent security (IDS) games on networks is notoriously slow, especially when downstream tasks like auditing or incentive design require re-solving the game under many parameter changes. Researchers (Elnaz Nowrouzi, Zhiqun Zuo, Xueru Zhang, Mohammad Mahdi Khalili) introduce BRAID (Best-Response Amortized Iterative Dynamics), a graph neural network that learns a direct map from game parameters to equilibrium effort profiles. Instead of running hundreds of best-response iterations, BRAID makes a single forward pass, achieving up to 43x faster per instance across utility specs including log-linear, quadratic-cost, and log CES.

BRAID's architecture draws directly from the best-response fixed-point structure: SUM aggregation captures additive neighbor coupling, while a weight-tied GRU mimics a damped best-response update. This design lets the same model generalize across different cost curvatures and network topologies. Beyond equilibrium prediction, BRAID recovers how equilibrium efforts shift under perturbations to costs and edge weights. The authors introduce two training strategies—interior-equilibrium training and input-noise regularization—that improve the model's local sensitivity behavior without needing sensitivity labels. Experiments confirm accurate equilibrium prediction and sensitivity recovery across network sizes, making BRAID a practical accelerator for security game analysis.