Sequential Distribution-Shift Detection — DI-SCUSUM

Co-developed DI-SCUSUM, a training-free streaming detector using closed-form Gaussian-smoothed scores and KL-based detection-delay and false-alarm analysis under empirical models. Evaluations covered Gaussian simulations, MNIST, and Oxford-IIIT Pet. In a calibrated anisotropic Gaussian experiment at matched false-alarm levels, the method achieved approximately 91% lower delay than score-based CUSUM.

Equal contribution: Arman Adibi and Mohammadreza Maleki. Collaborators: Sanjeev Kulkarni and H. Vincent Poor at Princeton University.

See Research and Selected Projects for the project summary.