By Jayin Khanna

The Github Repo:

https://github.com/ddgoodgood/TRCE

Goal: Concept Erasure

Modify the model weights in a way so that it unlearns to generate specific concepts anymore

This is done in 2 stages:

→ Stage 1: Textual Semantic Erasure (TSE) This is done to remove the malicious content from the word embeddings

→ Stage 2: Denoising Trajectory Steering (DTS) Even after removing the malicious content from the word embeddings, it can so happen that the model still generates malicious content using ‘non malicious’ words, and the diffusion trajectories are already biased towards it Hence, we need to unlearn the diffusion trajectories that lead to ‘malicious’ content generation in the initial steps of diffusion denoising

Stage 1 Architecture and specifics

Stage 2 Architecture and specifics

Notation I used for Architecture Diagrams

Stage 1 Architecture Diagram— From Claude

Stage 2 Architecture Diagram— From Claude