Evading-DeepFake-Detectors-via-Conditional-Diffusion-Models

This repository provides the official PyTorch implementation of our paper:

Evading DeepFake Detectors via Conditional Diffusion Models
IH&MMSec 2024, ACM Workshop on Information Hiding and Multimedia Security
[Paper (PDF)]

🧠 Overview

DeepFake detectors play a critical role in identifying manipulated facial images. However, they are vulnerable to adversarial attacks. This work proposes a semantic reconstruction-based adversarial attack that leverages conditional diffusion models to generate imperceptible perturbations in the latent space.

Key Highlights

• 🌀 Latent space attack using DDIM-guided conditional sampling.
• 🕵️ Evades both spatial and frequency domain detectors (e.g., XceptionNet, SRM).
• 🖼️ High visual quality — minimal artifacts, low LPIPS, and low FID.
• 🎯 Supports white-box and black-box settings with meta-learning transferability.

---

🔧 Method

The proposed attack optimizes the latent representation of fake images in a semantic latent space extracted by a diffusion autoencoder. A conditional DDIM decoder reconstructs adversarial faces guided by:

• Attack loss: Cross-entropy to mislead the detector.
• Perceptual loss: LPIPS + L2 to preserve structure and realism.