New AI Research: How Human Perception is Influenced by Adversarial Images
New research demonstrates that even small changes to digital images, meant to confuse computer vision systems, can also impact human perception. It’s important to understand how AI and human vision differ when it comes to interpreting visual inputs.
What Are Adversarial Images?
Adversarial images are digital images that have been subtly altered to make an AI model incorrectly classify the contents with high confidence. Attacks can be targeted in several ways, such as making a model see a vase as a cat, or they can be designed to make the misclassification as anything except a vase.
How Do Adversarial Examples Influence Human Perception?
The research shows that humans are indeed influenced by adversarial perturbations in images, even those that are insignificant and would not typically be noticed. Humans struggle with distinguishing the slight differences between original and adversarial images and often make arbitrary choices when prompted to identify the differences between them.
The Importance of AI Safety and Security Research
Understanding how adversarial images affect human perception is crucial for AI safety and security research. These findings can inform future research seeking to improve the robustness of computer vision models by aligning them better with human visual representations. It also highlights the importance of cognitive science and neuroscience to better understand AI systems and their potential impacts as we work to build safer, more secure systems. This continued research will help ensure that AI is safe and beneficial for all.
To learn more about AI and human perception, click here.