DVAR: Adversarial Multi-Agent Debate for Video Authenticity Detection

1. Problem Statement

The rapid advancement of AI video generation (e.g., diffusion models) has created highly realistic synthetic content that is increasingly difficult to distinguish from real footage.

• Limitations of Current Methods: Existing deepfake detection approaches rely on supervised learning and pattern matching (learning specific artifact representations from labeled datasets).
  • Generalization Failure: These models often fail when encountering unseen generative architectures (Out-of-Distribution or OOD scenarios).
  • Lack of Interpretability: They function as "black boxes," relying on opaque correlations rather than explicit physical reasoning, often overlooking blatant physical violations.
  • LVLM Limitations: While Large Vision-Language Models (LVLMs) show promise, they suffer from reasoning hallucinations, a granularity gap (ignoring subtle cues), and hypothesis ambiguity when applied directly to detection tasks.

Goal: To develop a training-free, interpretable, and generalizable framework for video authenticity detection that moves beyond pattern matching to structured forensic reasoning.

2. Methodology: The DVAR Framework

DVAR (Debate-based Video Authenticity Reasoning) reformulates video detection as a structured, four-stage forensic reasoning process. It does not require task-specific training but leverages an external knowledge base and multi-agent debate.

Core Components

1. GenVideoKB (Knowledge Repository):

   • A dynamic, external knowledge base containing high-level reasoning heuristics about generative boundaries, synthesis mechanisms, and failure modes.
   • It provides positive guidance (plausible generative explanations) and negative guidance (common reasoning pitfalls) to support evidence interpretation without encoding specific dataset features.

2. Stage 1: Evidence Discovery

   • The system samples frames from the input video to create a clip representation.
   • A multimodal perception module analyzes the scene to generate a Scene Observation and identifies Forgery Traces ($T$).
   • Traces are observable anomalies (e.g., temporal jitter, inconsistent lighting) that deviate from physical expectations, supported by priors from GenVideoKB.

3. Stage 2: Multi-Agent Adversarial Debate

   • For each identified trace, two agents engage in a multi-round debate to explain the anomaly:
     • Generative Hypothesis Agent (GHA): Argues the trace is a synthetic artifact (e.g., diffusion instability, VAE compression).
     • Natural Mechanism Agent (NMA): Argues the trace is a result of natural physical processes (e.g., wind, rolling shutter, depth of field).
   • Mechanism: Agents must actively propose hypotheses, contest opposing views with contradictory evidence, and rebut challenges. This process suppresses hallucinations and forces logical convergence.
   • Outcome: The debate yields a structured record of competing hypotheses ($H_{gen}, H_{nat}$) and their auxiliary assumptions.

4. Stage 3: Explanatory Cost Calculation (Adjudication)

   • The system applies Occam's Razor via the Minimum Description Length (MDL) framework.
   • Explanatory Cost: The "logical burden" of a hypothesis is quantified by the token length of its compressed explanation. A hypothesis requiring fewer auxiliary assumptions to explain the evidence is considered more plausible.
   • Cost Gap ($\Delta \ell$): The difference in description length between the natural and generative hypotheses determines the winner for unresolved traces.
   • Resolved traces (where one agent concedes) provide categorical signals ($-1$ or $+1$).

5. Stage 4: Final Decision Arbiter

   • An Arbiter module aggregates signals from all traces (both resolved decisions and cost gaps).
   • It consolidates heterogeneous evidence into a unified verdict (REAL/FAKE), confidence score, and a structured rationale, operating under a fixed prompt template to ensure reproducibility.

3. Key Contributions

1. DVAR Framework: A novel, training-free paradigm that transforms video detection into a structured forensic reasoning problem. It eliminates reliance on task-specific datasets by using adversarial debate and MDL-based cost comparison.
2. GenVideoKB: The introduction of a dynamic knowledge repository that provides semantic-level priors on generative mechanisms. This mitigates reasoning hallucinations and enhances stability in open-world scenarios without causing data leakage.
3. Interpretability & Robustness: The framework provides explicit reasoning traces (debate logs and cost analysis), making the detection process transparent. It achieves robustness against unseen generative architectures by focusing on physical inconsistencies rather than learned artifacts.

4. Experimental Results

The authors evaluated DVAR on two major benchmarks: Buster (focusing on OOD generalization) and GenVidBench (large-scale, diverse generators).

• OOD Generalization (Buster Benchmark):

  • DVAR achieved 92.2% Accuracy and 94.7% F1 Score.
  • This performance is comparable to the state-of-the-art supervised method BusterX++ (92.4% ACC), despite DVAR having no training data from the benchmark.
  • Supervised baselines (e.g., 3D ResNet, ViViT) showed significant performance drops on unseen generators, whereas DVAR remained stable.

• Large-Scale Performance (GenVidBench):

  • On a stratified subset of GenVidBench, DVAR achieved 95.52% Accuracy and 98.22% F1 Score, outperforming the supervised DeMamba model (89.06% ACC) evaluated under the same setting.

• Ablation Studies:

  • Structured Reasoning: Direct prompting of LVLMs yielded only ~52% accuracy. Adding Evidence Discovery and Debate stages progressively improved performance, confirming the necessity of the structured pipeline.
  • GenVideoKB: Integrating the knowledge base provided a significant boost (+5.01% ACC), proving that general LVLM knowledge is insufficient without domain-specific priors.
  • Backbone Robustness: DVAR maintained high performance across different LVLM backbones (GPT-4o, GPT-5.4, Qwen-3.5-Plus), demonstrating model-agnostic effectiveness.

• Efficiency: While DVAR incurs higher token consumption (~3x direct prompting) due to multi-stage reasoning, it avoids the massive computational costs of training large supervised models.

5. Significance

• Paradigm Shift: DVAR marks a shift from data-driven pattern matching to reasoning-driven forensics. It addresses the fundamental limitation of supervised models: their inability to generalize to unseen generation technologies.
• Open-World Applicability: By being training-free and relying on physical reasoning rather than dataset-specific artifacts, DVAR offers a sustainable solution for detecting AI-generated content in rapidly evolving open-world scenarios.
• Transparency: Unlike black-box classifiers, DVAR provides human-readable debate logs and cost analyses, offering explicit justifications for its decisions, which is crucial for trust and accountability in media forensics.
• Scalability: The framework can be updated via the GenVideoKB without retraining the core detection model, allowing it to adapt to new generative models through knowledge injection rather than data re-collection.