MMA-Diffusion: MultiModal Attack on Diffusion Models

TL;DR

MMA-Diffusion demonstrates a novel multi-modal attack method that effectively bypasses safety measures in Text-to-Image models, revealing significant security vulnerabilities in current defenses for both open-source and commercial systems.

Contribution

The paper introduces MMA-Diffusion, a multi-modal attack framework that circumvents existing safety mechanisms in T2I models, exposing critical security flaws.

Findings

Successfully bypasses prompt filters and safety checkers

Reveals vulnerabilities in open-source and commercial T2I models

Highlights need for improved safety defenses

Abstract

In recent years, Text-to-Image (T2I) models have seen remarkable advancements, gaining widespread adoption. However, this progress has inadvertently opened avenues for potential misuse, particularly in generating inappropriate or Not-Safe-For-Work (NSFW) content. Our work introduces MMA-Diffusion, a framework that presents a significant and realistic threat to the security of T2I models by effectively circumventing current defensive measures in both open-source models and commercial online services. Unlike previous approaches, MMA-Diffusion leverages both textual and visual modalities to bypass safeguards like prompt filters and post-hoc safety checkers, thus exposing and highlighting the vulnerabilities in existing defense mechanisms.