Stochastic Safe-Set Projection

We propose \emph{Stochastic Safe-Set Projection (SSP)}, a theoretically grounded and computationally efficient method for enforcing constraints in deep learning models. SSP performs a minibatch-level stochastic projection of model outputs onto a linearized feasible set estimated from the same batch, ensuring constraint satisfaction \emph{in expectation}. We provide formal guarantees on constraint violation, discuss extensions to multiple and nonlinear constraints, and validate SSP on synthetic toy problems that illustrate its effectiveness for bounded outputs, fairness, and linear separability. Our results demonstrate that SSP is scalable, simple to implement, and robust across different settings, making it suitable for safety-critical and trustworthy AI applications.