Natural Adversarial Sample

Natural adversarial samples are inputs subtly altered to deceive machine learning models without noticeably changing their appearance to humans. Current research focuses on generating these samples, particularly for image and text data, using methods like evolutionary algorithms and probabilistic labeling to create realistic, yet adversarial, datasets. This work is crucial for evaluating and improving the robustness of AI systems across various applications, including healthcare and safety-critical systems, by moving beyond reliance on artificially generated adversarial examples that may not reflect real-world threats. The ultimate goal is to develop more resilient models capable of handling naturally occurring adversarial inputs.