Sealing The Backdoor

This repository contains the code and a link to model weights results from the work presented in the paper Sealing The Backdoor: Unlearning Adversarial Text Triggers In Diffusion Models Using Knowledge Distillation - Arxiv

The code files are:

1. self_kd.py - Self-Knowledge Distillation
2. attention_guided_kd.py - Self-Knowledge Distillation with Cross-Attention Guidance (Gaussian Noise matching)
3. attention_guided_kd_black.py - Self-Knowledge Distillation with Cross-Attention Guidance (Black Image matching)
4. attention_guided_kd_random_words.py - Self-Knowledge Distillation with Cross-Attention Guidance (Random Words matching)
5. finetune_rev.py - Finetune reversal of poisoning

The attention capture mechanism (in attention_map folder) is adapted from https://github.com/wooyeolBaek/attention-map

Model weights before and after unpoisoning can be found in this Huggingface Repo