CLIP computer vision system fooled by handwritten notes

OpenAI researchers discovered that their CLIP computer vision system could be deceived by tools as simple as a pen and paper. The finding highlighted how easily the system could be fooled.

Launched January 2023, CLIP was billed as a state-of-the-art system that explores how AI systems can learn to identify objects without close supervision by training on a database comprising 400 million image-text pairs scraped from the internet. 

The OpenAI study discovered CLIP includes so-called 'multimodal neurons' which respond to an abstract concept as a word or as a picture, thereby making it more powerful. However, they also potentially opened the system to abuse, such as adding dollar signs over a photograph of a poodle to fool the system into thinking the dog was a piggy bank.

News, commentary, analysis 🗞️

• https://www.theverge.com/2021/3/8/22319173/openai-machine-vision-adversarial-typographic-attacka-clip-multimodal-neuron

• https://www.sciencetimes.com/articles/30057/20210309/openai-deceived-handwritten-notes-fool-computer-vision-typographic-attack.htm

• https://boingboing.net/2021/03/09/openai-discovers-its-visual-ai-can-be-fooled-by-written-notes.html

• https://slashdot.org/story/21/03/08/2147254/openais-state-of-the-art-machine-vision-ai-fooled-by-handwritten-notes