RIML Lab

💠 Compositional Learning Journal Club

Join us this week for an in-depth discussion on Compositional Learning in the context of cutting-edge text-to-image generative models. We will explore recent breakthroughs and challenges, focusing on how these models handle compositional tasks and where improvements can be made.

✅ This Week's Presentation:

🔹 Title: Backdooring Bias into Text-to-Image Models

🔸 Presenter: Mehrdad Aksari Mahabadi

🌀 Abstract:
This paper investigates the misuse of text-conditional diffusion models, particularly text-to-image models, which create visually appealing images based on user descriptions. While these images generally represent harmless concepts, they can be manipulated for harmful purposes like propaganda. The authors show that adversaries can introduce biases through backdoor attacks, affecting even well-meaning users. Despite users verifying image-text alignment, the attack remains hidden by preserving the text's semantic content while altering other image features to embed biases, amplifying them by 4-8 times. The study reveals that current generative models make such attacks cost-effective and feasible, with costs ranging from 12 to 18 units. Various triggers, objectives, and biases are evaluated, with discussions on mitigations and future research directions.

📄 Paper: Backdooring Bias into Text-to-Image Models

Session Details:
- 📅 Date: Sunday
- 🕒 Time: 5:00 - 6:00 PM
- 🌐 Location: Online at vc.sharif.edu/ch/rohban


We look forward to your participation! ✌️

arXiv.org

Backdooring Bias into Text-to-Image Models

Text-conditional diffusion models, i.e. text-to-image, produce eye-catching images that represent descriptions given by a user. These images often depict benign concepts but could also carry other...

www.tgoop.com/RIMLLab/140

2.8K viewsAmir Kasaei, Oct 26 at 15:23

💠 Compositional Learning Journal Club

Join us this week for an in-depth discussion on Compositional Learning in the context of cutting-edge text-to-image generative models. We will explore recent breakthroughs and challenges, focusing on how these models handle compositional tasks and where improvements can be made.

✅ This Week's Presentation:

🔹 Title: Backdooring Bias into Text-to-Image Models

🔸 Presenter: Mehrdad Aksari Mahabadi

🌀 Abstract:
This paper investigates the misuse of text-conditional diffusion models, particularly text-to-image models, which create visually appealing images based on user descriptions. While these images generally represent harmless concepts, they can be manipulated for harmful purposes like propaganda. The authors show that adversaries can introduce biases through backdoor attacks, affecting even well-meaning users. Despite users verifying image-text alignment, the attack remains hidden by preserving the text's semantic content while altering other image features to embed biases, amplifying them by 4-8 times. The study reveals that current generative models make such attacks cost-effective and feasible, with costs ranging from 12 to 18 units. Various triggers, objectives, and biases are evaluated, with discussions on mitigations and future research directions.

📄 Paper: Backdooring Bias into Text-to-Image Models

Session Details:
- 📅 Date: Sunday
- 🕒 Time: 5:00 - 6:00 PM
- 🌐 Location: Online at vc.sharif.edu/ch/rohban


We look forward to your participation! ✌️

BY RIML Lab