Rzl-Dzl-AI

At the exhibition from January 10, 2018 to March 18, 2018

»Rzl-Dzl-AI« demonstrates how deep neural networks are easily fooled, a dystopian reality when you realize that, for example, the military is already using them. The question is how much confidence do we have in ourselves and the technologies we develop?

»Rzl-Dzl-AI« consists of encoded neural network-generated images that mislead and hack other neural networks or AI-based image recognition systems. The encoded images and patterns were generated through evolutionary algorithms. The images are heavily abstracted and unrecognizable to humans, but state-of-the-art, cutting edge convolutional neural networks (deep neural networks trained on ImageNet) believe them to be a familiar object with ≥ 99.99 % certainty.

»Rzl-Dzl-AI« is the next level »razzle-dazzle«, an old expression used to describe the effect of a new camouflage technique used extensively in World War I, and to a lesser extent in World War II; afterwards, it reduced collateral damage by more then 10 %. The camouflage consists of complex patterns of geometric shapes in contrasting colors, interrupting and intersecting each other. Unlike other forms of camouflage, the intention of razzle-dazzle is not to conceal, but to make it difficult to estimate a target’s range, speed, and direction.