back

Marcel Top

Poison Data, Kill Algorithms

Poison Data, Kill Algorithms examines developments in surveillance technology, looking closely at the way private surveillance companies buy large quantities of data to train facial recognition algorithms. This data is often supplied by poorly paid workers who photograph themselves wearing silicone masks with various characteristics. The images generated are deliberately low quality, in order to train algorithms to handle the kind of imagery they are likely to encounter in real-world use.

Marcel Top Poison Data, Kill Algorithms 2025

Artist Marcel Top researched these technologies and decided to create his own dataset — with a hidden purpose. Top photographed himself wearing over 1,000 combinations of 11 commonly used protest disguises, worn on top of a lifelike silicone mask. As algorithms are also trained to detect suspicious movements, Top recorded a series of movements across 500 images and 5 videos: walking, walking whilst holding a sign, sprinting, and throwing. The dataset is composed in such a way that it appears highly attractive to surveillance companies: freely available and ostensibly well-suited to improving algorithms that identify protesters.

What distinguishes the dataset is the presence of poisoned data: in 10% of the images, a small red pixel is concealed, combined with deliberately incorrect annotations. Algorithms trained on this data become corrupted and will misinterpret patterns beyond the dataset, which renders them unreliable. Poison Data, Kill Algorithms inverts automated data collection, offering a tool of resistance and resilience to those who are tracked by those very same systems.

Location: Carré Chassé

More information