Universal Adversarial

Universal adversarial attacks exploit subtle input manipulations to deceive machine learning models, particularly deep neural networks, across various modalities including images, text, and LiDAR data. Current research focuses on developing both effective attacks (e.g., using gradient-based methods, generative models, and optimized noise) and robust defenses (e.g., employing randomized smoothing, adversarial training, and diffusion models), often within black-box settings to address privacy concerns. This field is crucial for evaluating and improving the reliability and security of AI systems deployed in high-stakes applications like autonomous driving, healthcare, and language processing, where model robustness is paramount.