Semantics Preserved Distortion

Semantics-preserved distortion focuses on modifying data, such as images or text, to hinder unwanted inferences while retaining essential semantic information. Current research explores methods like meta-learning for robust image perturbation and regression-free approaches for image quality assessment that leverage semantic and distortion features. This field is crucial for privacy protection in machine learning applications, enabling the development of models trained on sensitive data without compromising individual privacy, and improving the robustness of image and text processing systems to various distortions.