Code for DPAgg-TI Experiments

We do not have a centralized codebase as we mainly built on existing implementations, namely

• The HuggingFace diffusers libray for Textual Inversion (link)
• The official Universal Guidance implementation for style guidance (link)

Please refer to the respective codebases for installation and usage details.

We provide the following 4 files:

• dpagg-ti.py - implementation for the embedding aggregation and image generation for DPAgg-TI
• train_dpagg-ti.sh - sample training script to train individual DPAgg-TI embeddings (run in /diffusers/examples/textual_inversion/)
• textual_inversion_dp-sgd.py - integrates Opacus DP-SGD into the HuggingFace implementation
• dpagg-style_guidance.py - implementation for the embedding aggregation and image generation for our approach with style guidance (run in /Universal-Guided-Diffusion/stable-diffusion-guided/scripts)

To respect the intellectual properties of @eveismyname and the International Olympic Committee, we do not provide our data. We appreciate your understanding.