Turns Out I’m Not Real: Towards Robust Detection of AI-Generated Videos

🌈 Abstract

The article discusses the development of generative models for creating high-quality videos, which has raised concerns about digital integrity and privacy vulnerabilities. It proposes a novel framework called DIVID (DIffusion-generated VIdeo Detector) for detecting videos synthesized from multiple state-of-the-art (SOTA) generative models, such as Stable Video Diffusion. The key points are:

• SOTA methods for detecting diffusion-generated images lack robustness when applied to diffusion-generated videos, as they struggle to capture temporal features and dynamic variations between frames.
• The article introduces a new benchmark video dataset for diffusion-generated videos using SOTA video creation tools.
• DIVID extracts representations from the diffusion model for video frames and trains a CNN + LSTM architecture to capture temporal features, achieving 93.7% detection accuracy for in-domain videos and improving accuracy for out-domain videos by up to 16 points.

🙋 Q&A

[01] Introduction

1. What are the key concerns raised by the impressive achievements of generative models in creating high-quality videos?

• The impressive achievements of generative models in creating high-quality videos have raised concerns about digital integrity and privacy vulnerabilities.

2. What is the focus of recent works to combat Deepfakes videos?

• Recent works have developed detectors that are highly accurate at identifying GAN-generated samples, but the robustness of these detectors on diffusion-generated videos generated from video creation tools is still unexplored.

3. What is the proposed solution in this paper?

• The paper proposes a novel framework called DIVID (DIffusion-generated VIdeo Detector) for detecting videos synthesized from multiple state-of-the-art (SOTA) generative models, such as Stable Video Diffusion.

[02] Related Works

1. What are the key developments in diffusion-based video generation?

• Diffusion-based video generation tools like SORA by OpenAI, Stable Video Diffusion, MidJourney, RunwayML, Show-1, Pika, and Deep Dream Generator enable users to generate videos with impressive visual and narrative quality, showcasing a range of capabilities from enhancing video quality to generating entirely new content.

2. What are the key challenges in detecting deepfake videos?

• Traditional DNN networks or audio-visual approaches based on lipsync inconsistency detection are not robust enough to detect deepfake videos. Prior works have proposed techniques like tracking and extracting facial information, or using CNN-based models, but they were found ineffective for generalizing to diffusion-based images.

3. What are the key limitations of prior works on detecting diffusion-generated images?

• While prior works like DIRE and SeDID have shown success in detecting diffusion-generated images, the robustness of these detectors on diffusion-generated videos remains unexplored.

[03] Method

1. What is the key idea behind the Denoising Diffusion Probabilistic Models (DDPM)?

• DDPM operates by gradually transforming random noise into structured images over a series of steps, simulating a reverse diffusion process.

2. How does the Denoising Diffusion Implicit Models (DDIM) differ from DDPM?

• DDIM modifies the diffusion process by introducing a non-Markovian implicit trajectory, which can train the diffusion more efficiently without content distortion.

3. How does DIVID leverage the Diffusion Reconstruction Error (DIRE) to detect diffusion-generated videos?

• DIVID leverages the DIRE, which measures the difference between an input image and its corresponding reconstructed version from the diffusion model, to detect diffusion-generated videos. It uses a CNN+LSTM architecture to capture the temporal features of RGB frames and DIRE values.

[04] Experiment

1. What are the key components of the dataset used in the experiments?

• The dataset includes in-domain video clips generated using Stable Video Diffusion, and out-domain video clips generated using Pika, Gen-2, and SORA.

2. How does DIVID's performance compare to the baselines on the in-domain and out-domain testsets?

• On the in-domain testset, DIVID achieves 98.20% average precision and outperforms the baselines by 0.94% to 3.52%. On the out-domain testsets, DIVID improves the average accuracy by 0.69% to 16.1% compared to the baselines.

3. What is the impact of different diffusion steps and DDIM steps on the performance of DIVID?

• The evaluation shows that the choice of diffusion steps and DDIM steps for generating the DIRE values has a significant impact on the performance of DIVID.