1 of 13

Data Attribution for Text-to-Image Models by Unlearning Synthesized Images

Richard Zhang

Jun-Yan Zhu

Alexei A. Efros

Sheng-Yu Wang

In NeurIPS, 2024.

Aaron Hertzmann

2 of 13

Influence scores

GenAI image

Data Attribution

Dataset

Stable Diffusion

Challenge: ground truth influence is unknown

Must intervene in the training process

3 of 13

Random subsets

Training�dataset

 

Counterfactual subset 1

Analyze models

 

Counterfactual subset 2N

Training 2N models is too expensive

 

c.f. Feldman & Zhang. What Neural Networks Memorize and Why. NeurIPS 2020.

Synthesized

4 of 13

Leave-one-out

 

 

Training�dataset

Unlearning

 

 

Koh & Liang. ICML 2017; Schioppa AAAI 2022; Park ICML 2023; Georgiev ICML Wkshp 2023; Grosse ArXiv 2023.

Evaluate

Influence functions: linear approx.�for unlearning & evaluation

Synthesized

 

Store a low-dimensional version�or recompute at test-time

5 of 13

Leave-one-out

Training�dataset

 

Synthesized

Can we remove the linear approximations?

Storing or recomputing�N models intractable

 

 

Unlearning

 

Evaluate

Koh & Liang. ICML 2017; Schioppa AAAI 2022; Park ICML 2023; Georgiev ICML Wkshp 2023; Grosse ArXiv 2023.

6 of 13

Attribution by Unlearning (AbU)

Assess influence

(by loss increase)

Training�dataset

Unlearning

 

 

Synthesized

7 of 13

Counterfactual evaluation

Training�dataset

Remove Top-K�Influential Images

 

Counterfactual subset

c.f. K. Georgiev, et al. How Training Data Guides Diffusion Models. In ArXiv, 2023.

Synthesized

Influences

8 of 13

Counterfactual evaluation

Failed Resynthesis

Training�dataset

Remove Top-K�Influential Images

 

Counterfactual subset

 

2. Regenerate & Assess difference

1. Check DDPM loss

If critical training images are identified,�removing them should destroy the generation

Expensive evaluation…

…but let’s do it! (for modest sizes)

c.f. K. Georgiev, et al. How Training Data Guides Diffusion Models. In ArXiv, 2023.

Synthesized

Influences

9 of 13

MS-COCO results

Ours

“A bus traveling on a freeway next to other traffic.”

DINO

JourneyTRAK

c.f. K. Georgiev, et al. How Training Data Guides Diffusion Models. In ArXiv, 2023.

Remove K=500

(0.4% of dataset)

Ours

DINO

JourneyTRAK

Attribution results

D-TRAK

D-TRAK

c.f. X. Zheng, et al. Intriguing Properties of Data Attribution on Diffusion Models. In ICLR, 2024.

Effective removal

10 of 13

Nearest neighbors

Tuned via Customization

Influence functions

11 of 13

Local attribution

“A motorcycle and a stop sign.”

Cropped

Queries

Attributed training images

12 of 13

Customized Model Benchmark

Object Centric

Generated Sample

DINO (AbC)

Ours

CLIP (AbC)

“An ink drawing of�V* turtle”

“A picture of tree in the style of V* art”

Ours

DINO (AbC)

CLIP (AbC)

Ours

D-TRAK

D-TRAK

13 of 13

Thank you!