Researchers at Columbia University's School of Engineering and Applied Science have introduced a new tool, DIVID, designed to identify AI-generated videos with 93.7% accuracy.
DIVID: A new tool for detecting AI-generated videos
In response to the increasing sophistication of AI-generated content, a research team led by computer science professor Junfeng Yang developed DIVID, which stands for DIffusion-generated VIdeo Detector.
AI-generated videos are becoming increasingly realistic, making it difficult for both human observers and existing detection systems to distinguish real footage from artificially created content.
The researchers say that unlike previous AI models such as generative adversarial networks (GANs), which were detectable through visible anomalies such as pixel irregularities or unnatural movement, new AI techniques such as diffusion models generate high-fidelity videos, making them extremely difficult to distinguish from real footage.
Rather than delving into the workings of AI models like GPT-4 or Gemini, DIVID builds on the team's previous work with Raidar, a tool designed to detect AI-generated text by analyzing language patterns.
Raidar's approach measures the number of changes required to transform text: fewer edits indicate a higher likelihood of machine generation, as AI tends to produce consistent text with minimal corrections.
Related article: Luma's AI video generator, the dream machine, is flooded with users, surpassing OpenAI's Sora
DIRE Technique
Applying similar principles, DIVID scrutinizes diffusion-generated videos using DIRE (DIffusion Reconstruction Error) technique, which evaluates the discrepancy between an input video and a reconstruction using a pre-trained diffusion model to flag videos that may have been generated by AI.
By focusing on the inherent differences between AI-generated and real-world videos, DIVID aims to power the detection capabilities needed to combat the proliferation of deceptive visual content.
Yang, co-director of the Software Systems Lab, noted that Raidar's insights can be adapted from text to visual media, emphasizing their universal applicability: As AI-generated videos become increasingly realistic, the team aimed to leverage Raidar's insights to develop a tool that can accurately detect AI-generated videos.
Advances in AI technology, particularly video synthesis with diffusion models such as OpenAI's Sora and Runway Gen-2, underscore the urgency of robust detection mechanisms like DIVID.
These models progressively refine each video frame from random noise, achieving unprecedented realism and challenging traditional detection methods that rely on surface-level anomalies.
Through DIVID, Columbia researchers' ultimate goal is to mitigate the risks posed by AI-generated video in a variety of contexts, including preventing fraud and maintaining the integrity of digital content.
“Raidar's insight that AI output is often judged to be of higher quality by other AIs and therefore requires less editing is extremely powerful and goes beyond just text,” Yang said in a statement.
“Given that AI-generated videos are becoming increasingly realistic, we wanted to leverage Raidar's insights to create a tool that can accurately detect AI-generated videos,” he added.
The research team's findings were published in arXiv.
Related article: MIT unveils new algorithm that can learn languages just by watching videos
ⓒ 2024 TECHTIMES.com All rights reserved. Please do not reproduce without permission.
