Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024

2025-03-04

AI Summary

Rapid advancements in deepfake generation have outpaced mainstream academic detection benchmarks, leading to overly optimistic estimates of model generalization in real-world settings. Method: We introduce RealDeep24—the first 2024 real-world multimodal deepfake benchmark—comprising 44 hours of video, 56.5 hours of audio, and 1,975 images, collected from 88 websites and 52 languages across social and detection platforms. We propose an in-the-wild dynamic evaluation paradigm integrating cross-lingual/cross-platform provenance verification and an AUC-driven robustness assessment framework. Contribution/Results: Experiments reveal drastic performance drops for current SOTA detectors: AUC declines by 50% (video), 48% (audio), and 45% (image) relative to prior benchmarks. While commercial and fine-tuned models show notable improvements, they still underperform human experts. RealDeep24 and its evaluation toolkit are publicly released to foster realistic, rigorous deepfake detection research.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Computer Vision: Multi-modal Vision

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Security and Privacy: Large-scale security measurements

Abstract

In the age of increasingly realistic generative AI, robust deepfake detection is essential for mitigating fraud and disinformation. While many deepfake detectors report high accuracy on academic datasets, we show that these academic benchmarks are out of date and not representative of recent deepfakes. We introduce Deepfake-Eval-2024, a new deepfake detection benchmark consisting of in-the-wild deepfakes collected from social media and deepfake detection platform users in 2024. Deepfake-Eval-2024 consists of 44 hours of videos, 56.5 hours of audio, and 1,975 images, encompassing the latest manipulation technologies. The benchmark contains diverse media content from 88 different websites in 52 different languages. We find that the performance of open-source state-of-the-art deepfake detection models drops precipitously when evaluated on Deepfake-Eval-2024, with AUC decreasing by 50% for video, 48% for audio, and 45% for image models compared to previous benchmarks. We also evaluate commercial deepfake detection models and models finetuned on Deepfake-Eval-2024, and find that they have superior performance to off-the-shelf open-source models, but they do not yet reach the accuracy of human deepfake forensic analysts. The dataset is available at https://github.com/nuriachandra/Deepfake-Eval-2024.

Problem

• Address outdated deepfake detection benchmarks with new realistic data.
• Evaluate deepfake detection models on diverse, real-world media content.
• Assess performance gap between AI models and human forensic analysts.

Innovation

• Introduces Deepfake-Eval-2024 benchmark
• Includes diverse multi-modal deepfake data
• Evaluates state-of-the-art detection models