Adversarial Perturbations

Adversarial perturbation research studies how small, structured, or physically realizable changes to inputs alter model predictions and behavior. Common approaches include gradient-based attacks such as FGSM and PGD, universal and transfer attacks, query-efficient black-box search, and attacks designed for text, audio, 3D data, or physical settings. Defenses include adversarial training, input purification, detection, and robustness certification, while protective perturbations also aim to preserve privacy or limit unauthorized use of images and videos. A central challenge is that success on fixed benchmarks may not predict performance after compression, resizing, environmental change, or closed-loop deployment, and stronger protection often trades off against visual quality, utility, or attack transferability.