📢 Join the Adversarial Nibbler crowdsourcing challenge

[Lora Aroyo (GMail)]

Hi everyone, 

Here is a great opportunity to do some prompt hacking for text-to-image models and help them get safer!

📢 We just launched the 🏆 #AdversarialNibbler challenge 🏆 at https://www.dataperf.org/adversarial-nibbler in collaboration with Kaggle, MLCommons and DataPerf. 

Adversarial Nibbler is crowdsourcing the discovery of implicitly adversarial prompts (i.e. innocuous "safe" looking prompts that still produce harmful "unsafe" images) that can help us improve the #safety of #text2image #GenerativeAI. The challenge is hosted on Dynabench platform and is designed to be accessible to a wide range of people with and without a traditional AI/ML background.

📜 After participating you also can submit a short paper with your experiences, red-teaming approaches, reflections and results to the "The Art of Safety" Workshop

Details about the challenge: Adversarial Nibbler: A Data-Centric Challenge for Improving the Safety of Text-to-Image Models (paper)
Join the challenge: https://dynabench.org/tasks/adversarial-nibbler (website)
Submit workshop paper: by August 25th, 2023 (11:59pm AoE)

Team behind the challenge and workshop: 

Alicia Parrish (Google), 

Lora Aroyo (Google), 

Hannah Rose Kirk (University of Oxford), 

Oana Inel (University of Zurich), 

Charvi Rastogi (CMU), 

Max Bartolo (Cohere),

Jessica Quaye (Harvard University), 

Vijay Janapa Reddi (Harvard University)

Please reach out if you are interested and would like to know more about it

Best

Lora

========================
Lora Aroyo
https://en.wikipedia.org/wiki/Lora_Aroyo 

http://lora-aroyo.org

Twitter: @laroyo

[Lora Aroyo]

Hi everyone, 

Here is a great opportunity to do some prompt hacking for text-to-image models and help them get safer!

📢 We just launched the 🏆 #AdversarialNibbler challenge 🏆 at https://www.dataperf.org/adversarial-nibbler in collaboration with Kaggle, MLCommons and DataPerf. 

Adversarial Nibbler is crowdsourcing the discovery of implicitly adversarial prompts (i.e. innocuous "safe" looking prompts that still produce harmful "unsafe" images) that can help us improve the #safety of #text2image #GenerativeAI. The challenge is hosted on Dynabench platform and is designed to be accessible to a wide range of people with and without a traditional AI/ML background.

📜 After participating you also can submit a short paper with your experiences, red-teaming approaches, reflections and results to the "The Art of Safety" Workshop

Details about the challenge: Adversarial Nibbler: A Data-Centric Challenge for Improving the Safety of Text-to-Image Models (paper)
Join the challenge: https://dynabench.org/tasks/adversarial-nibbler (website)
Submit workshop paper: by August 25th, 2023 (11:59pm AoE)

Team behind the challenge and workshop: 

Alicia Parrish (Google), 

Lora Aroyo (Google), 

Hannah Rose Kirk (University of Oxford), 

Oana Inel (ETH Zürich), 

Charvi Rastogi (CMU), 

Max Bartolo (Cohere),

Jessica Quaye (Harvard University), 

Vijay Janapa Reddi (Harvard University)

Please reach out if you are interested and would like to know more about it

Best

Lora

--

Lora Aroyo Research Scientist DEER team, go/deer Data Excellence for Responsible AI