AVFakeBench: A Comprehensive Audio-Video Forgery Detection Benchmark for AV-LMMs

1. Problem Statement

The rapid advancement of generative AI (e.g., Sora, Kling) has escalated the threat of Audio-Video (AV) forgeries beyond human-centric deepfakes to include diverse manipulations in complex natural scenes. Existing benchmarks suffer from three critical limitations:

• Limited Scope: They focus almost exclusively on human subjects (faces) and single manipulation types (either editing or synthesis), failing to cover general subjects (e.g., landscapes, animals, industrial scenes) or hybrid forgery types.
• Shallow Annotation: Most datasets provide only binary labels (Real vs. Fake), lacking fine-grained details, forgery type classification, or explanatory reasoning required for real-world forensic analysis.
• Inadequate Evaluation: Current benchmarks cannot assess the ability of models to localize specific forgery traces or explain why a clip is forged, which is crucial for trust and verification.

2. Methodology: AVFakeBench Construction

A. Dataset Composition

AVFakeBench is a comprehensive benchmark containing 3,000 audio-video clips and 12,000 question-answer (QA) pairs.

• Subjects: Covers both Human Subjects (1,500 clips, focusing on speech) and General Subjects (1,500 clips) across 11 real-world scenarios (e.g., natural landscapes, animals, sports, industry, science).
• Forgery Taxonomy: Defines three content states (Real, Edited, Synthesized) and combines them across audio and video modalities to form 7 distinct multi-modal forgery types (e.g., Real Audio + Edited Video, Synthesized Audio + Synthesized Video).
• Multi-Level Annotations: The dataset supports four hierarchical evaluation tasks:
  1. L1: Binary Judgment: Is the clip real or forged?
  2. L2: Forgery Type Classification: Which of the 7 forgery categories does it belong to?
  3. L3: Forgery Detail Selection: Multiple-choice selection of the specific evidence (e.g., "audio silence at 00:03").
  4. L4: Explanatory Reasoning: Open-ended generation of forensic evidence and rationale.

B. Multi-Stage Hybrid Forgery Framework

To generate high-fidelity, diverse forgeries, the authors propose a framework integrating proprietary models for planning and expert generative models for execution:

1. Planning Stage: A Large Multimodal Model (LMM) generates detailed spatio-temporal manipulation plans (e.g., dynamic scene descriptions for synthesis or specific editing instructions like "remove the boat").
2. Execution Stage:
   • Synthesis Branch: Uses Image-to-Video models (e.g., Kling, QingYing) and Video-to-Audio generators to create synchronized content based on the plan.
   • Editing Branch: Uses segmentation tools (SAM2) and video editing models to perform localized insertions, deletions, or replacements, followed by audio re-synthesis.
3. Combination Stage: Real and forged streams are recombined to instantiate the 7 specific forgery types.
4. Human Supervision: Critical stages involve human oversight to ensure semantic consistency, physical plausibility, and cross-modal synchronization.

C. Annotation Pipeline

To generate L3 and L4 annotations without prohibitive manual effort, the authors use an LMM-assisted annotation pipeline:

• The LMM is fed multi-modal context inputs: video frames, motion heatmaps, Mel spectrograms, and high-frequency zoom maps.
• It generates detailed L4 rationales, which are then distilled into L3 multiple-choice questions with distractors.
• All generated annotations undergo human verification for correctness, plausibility, and clarity.

3. Key Contributions

1. First Comprehensive AV Forgery Benchmark: AVFakeBench is the first dataset to simultaneously cover human and general subjects, 11 diverse scenarios, 7 multi-modal forgery types, and 4 levels of annotation depth.
2. Novel Forgery Framework: Introduces a multi-stage hybrid framework that decouples intent planning from execution, enabling scalable, high-fidelity, and semantically consistent forgery generation.
3. Comprehensive Evaluation: Establishes a multi-task evaluation framework that moves beyond simple detection to assess fine-grained perception and explanatory reasoning.

4. Experimental Results

The authors evaluated 11 Audio-Video Large Language Models (AV-LMMs) (5 open-source, 6 proprietary) and 2 expert detection models on AVFakeBench.

• Binary Judgment (L1): AV-LMMs show strong potential, outperforming traditional expert detectors (which rely heavily on face artifacts) and demonstrating robustness across both human and general subjects.
• Fine-Grained Perception (L2 & L3): Performance drops significantly (~70% drop) when tasks require identifying specific forgery types or subtle editing operations. Models struggle particularly with Edited forgeries (localized changes) compared to Synthesized ones.
• Explanatory Reasoning (L4): AV-LMMs exhibit poor reasoning capabilities. Even top-tier proprietary models (e.g., GPT-4o, Gemini-2.5-Pro) achieve low scores in providing accurate, grounded forensic explanations.
• Modality Bias: A clear vision-dominant pattern exists. Models perform significantly better on video-only forgeries but collapse on audio-only forgeries, indicating a lack of robust cross-modal alignment and audio perception.
• Model Bias: Many models exhibit a strong bias toward predicting "Real" or specific categories regardless of input, driven by pretraining priors rather than actual forensic evidence.

5. Significance and Future Outlook

• Benchmarking Standard: AVFakeBench fills a critical gap in the field by providing a rigorous, extensible testbed for evaluating the next generation of AV-LMMs.
• Revealing Limitations: The study highlights that while AV-LMMs are promising unified detectors, they currently lack the fine-grained perception and logical reasoning required for reliable real-world forensic analysis.
• Research Direction: The results suggest that future research must focus on improving cross-modal alignment, reducing modality bias (especially in audio), and enhancing the explainability and reasoning capabilities of multimodal models.

In conclusion, AVFakeBench serves as a foundational resource for advancing the scientific understanding of audio-visual manipulation and guiding the development of robust, perception-aware, and explainable forgery detection systems.