Noise Robustness

Noise robustness studies how models and signal-processing systems can keep working when inputs, training labels, or hardware computations are corrupted. Approaches include noise augmentation and noise-aware training, robust losses for mislabeled data, denoising and speech enhancement, representation regularization, test-time adaptation, and randomized smoothing for certified guarantees. Evaluation ranges from image-corruption benchmarks such as CIFAR and ImageNet variants to speech tests across signal-to-noise levels, noisy clinical recordings, and hardware with measured device variation; useful metrics can include accuracy, word error rate, false alarms, safety outcomes, and certified radii. A central challenge is that robustness to one noise type may not transfer to mixtures or real-world shifts, while methods can trade clean performance for noisy-input gains and fixed benchmarks or aggregate scores may conceal critical failures.