Adversarial Encoder-Driven Filter for Targeted Image Tagging: A Novel Approach to Visual Content Manipulation

Computer vision, driven by artificial intelligence, has become pervasive in diverse applications such as self-driving cars and law enforcement. However, the susceptibility of these systems to attacks has raised significant concerns among researchers. This paper addresses the vulnerability of image tagging algorithms, particularly focusing on misclassifications induced by autoencoders. We present experiments conducted on Amazon Rekognition, where we developed a specialized autoencoder to manipulate the latent space, forcing it to align with specific tags. By integrating this manipulated latent space with other images, we demonstrate the ability to increase the confidence of a specific tag on Amazon Rekognition, leading to more false positives of the chosen tag. Our study showcases a practical method to exploit Amazon’s Rekognition image tagging algorithm using a black box approach.