AV-Deepfake1M: A Large-Scale LLM-Driven Audio-Visual Deepfake Dataset

Abstract: The detection and localization of highly realistic deepfake audio-visual content are challenging even for the most advanced state-of-the-art methods. While most of the research efforts in this domain are focused on detecting high-quality deepfake images and videos, only a few works address the problem of the localization of small segments of audio-visual manipulations embedded in real videos. In this research, we emulate the process of such content generation and propose the AV-Deepfake1M dataset. The dataset contains content-driven (i) video manipulations, (ii) audio manipulations, and (iii) audio-visual manipulations for more than 2K subjects resulting in a total of more than 1M videos. The paper provides a thorough description of the proposed data generation pipeline accompanied by a rigorous analysis of the quality of the generated data. The comprehensive benchmark of the proposed dataset utilizing state-of-the-art deepfake detection and localization methods indicates a significant drop in performance compared to previous datasets. The proposed dataset will play a vital role in building the next-generation deepfake localization methods. The dataset and associated code are available at https://github.com/ControlNet/AV-Deepfake1M .

• The paper introduces a large-scale dataset of over 1 million deepfake videos spanning more than 2,000 subjects with varied audio, video, and audio-visual manipulations.
• The paper presents a novel data generation pipeline that leverages LLMs like ChatGPT for transcript manipulation and high-quality audio-visual synthesis.
• The paper benchmarks state-of-the-art detection methods on AV-Deepfake1M, revealing significant performance drops that highlight the complexity of realistic deepfake challenges.

An Expert Review of "AV-Deepfake1M: A Large-Scale LLM-Driven Audio-Visual Deepfake Dataset"

The presented work introduces the AV-Deepfake1M dataset, a comprehensive resource for research in the domain of detecting and localizing audio-visual deepfake content. The paper highlights the challenges faced by current detection methods in identifying realistic deepfake media, underscoring the need for expansive and diverse datasets to train and evaluate new approaches.

Key Contributions:

1. Dataset Scope and Composition: AV-Deepfake1M is a large-scale dataset consisting of over 1 million deepfake videos involving more than 2,000 distinct subjects. It distinguishes itself through the generation of realistic deepfake content via a content-driven approach that includes video-only, audio-only, and audio-visual manipulations.
2. Novel Data Generation Pipeline: The authors introduce a robust pipeline for generating deepfake content, leveraging advanced models such as ChatGPT for transcript manipulation. The pipeline involves stages of transcript alteration, high-quality audio generation, and the creation of corresponding video, resulting in highly realistic and challenging benchmark data.
3. Benchmarking and Analysis: A thorough benchmark of existing detection and localization methodologies is conducted using the AV-Deepfake1M dataset. The results show a significant decline in performance for top-performing models on prior datasets when evaluated with AV-Deepfake1M, indicating the higher complexity and realism of this data.
4. Quality Assurance and Evaluation: AV-Deepfake1M videos maintain visual and auditory quality, rated using metrics such as PSNR, SSIM, SECS, and more. The dataset's fine granularity in temporal and modality manipulations adds an extra layer of difficulty, instrumental for developing future-proof detection models.
5. Human Evaluation Studies: To ensure a high level of realism, the dataset was subjected to human assessment, demonstrating the challenges in manual detection of the manipulations incorporated.

Implications and Prospective Developments:

• Research Utility:

The dataset is poised to become a crucial benchmark for advancing deepfake detection capabilities, fostering innovation in methods capable of fine-grained and multimodal detection.

• Theoretical Investigations:

AV-Deepfake1M will stimulate theoretical research into the nature of media synthesis, facilitating the exploration of new adversarial and generative approaches, and providing insights into the failure points of current methods.

• Practical Applications:

By simulating real-world challenges, AV-Deepfake1M can significantly enhance the robustness of systems against misinformation and identify discrepancies in both audio and visual modalities.

In conclusion, the AV-Deepfake1M dataset constitutes a significant step towards equipping the research community with necessary tools to counter the evolving challenges within the field of deepfake detection. Future investigations will likely focus on leveraging this dataset to develop algorithms with superior generalization and accuracy in real-world applications. The comprehensive scale and variety of AV-Deepfake1M establish a new standard for datasets in this field, ensuring its relevance for forthcoming advancements in AI-driven media authentication.