TriDF: Evaluating Perception, Detection, and Hallucination for Interpretable DeepFake Detection

Conference: CVPR 2026
arXiv: 2512.10652
Code: https://j1anglin.github.io/TriDF/
Area: Hallucination Detection
Keywords: DeepFake Detection, Interpretable Detection, Multi-modal Large Language Models, Hallucination Evaluation, Artifact Taxonomy

TL;DR

This paper proposes TriDF, the first benchmark to comprehensively evaluate interpretable DeepFake detection across three dimensions: Perception, Detection, and Hallucination. Comprising 55K high-quality samples covering 16 DeepFake types and 3 modalities, it reveals the tripartite coupling relationship where accurate perception is the foundation of reliable detection, but hallucination severely undermines decision-making.

Background & Motivation

1. Background: With the rapid advancement of generative models, DeepFake detection has evolved from simple binary classification to a requirement for interpretability—judging not only authenticity but also providing reasons for the verdict. Multi-modal Large Language Models (MLLMs) are increasingly utilized for interpretable DeepFake detection.

2. Limitations of Prior Work:

   • Coarse-grained Annotations in Existing Datasets: FF++, DFDC, and others only provide binary labels, failing to evaluate interpretability.
   • Incomplete Benchmark Coverage: DD-VQA covers only 4 forgery types, FakeBench covers 1, and LOKI covers 3; most support only the image modality, lacking cross-modal coverage.
   • Lack of Hallucination Evaluation: MLLMs may generate "hallucinations" (citing non-existent artifacts) when providing explanations. This is particularly dangerous in DeepFake detection, as false justifications can mislead judgment. Existing benchmarks entirely overlook this aspect.
   • Reliance on MLLMs to Judge MLLMs: Many benchmarks use GPT-4o to judge outputs from other models, introducing self-preference bias.

3. Key Challenge: Interpretable DeepFake detection requires models to simultaneously possess three capabilities—perceiving artifacts, detecting correctly, and explaining reliably—yet no unified framework exists to evaluate these three and their interdependencies.

4. Goal: To construct a comprehensive benchmark for interpretable DeepFake detection that unifiedly evaluates perception, detection, and hallucination to reveal their coupling relationships.

5. Key Insight: Starting from a human-annotated fine-grained artifact taxonomy, the authors establish quantifiable perception evaluations; pair real and fake samples to support hallucination detection; and cover three modalities (image/video/audio) and 16 DeepFake types.

6. Core Idea: Perception-Detection-Hallucination forms an inseparable triad for interpretable DeepFake detection, and TriDF is the first unified benchmark to evaluate all three simultaneously.

Method

Overall Architecture

TriDF decomposes the evaluation of whether an MLLM can trustworthily explain a DeepFake into three measurable tasks: whether it can see artifacts (Perception), whether it can make correct binary judgments after seeing them (Detection), and whether it fabricates non-existent artifacts in its explanation (Hallucination). To support these measurements, the pipeline is divided into two phases: Data Construction, which involves collecting face materials from public datasets, generating paired real-fake samples using 16 DeepFake techniques, and performing manual annotation based on a fine-grained artifact taxonomy after automatic quality control; and Benchmark Evaluation, which packages each sample into three types of questions (True/False, Multiple Choice, and Open-ended) for MLLMs to answer, followed by scoring via specific metrics for perception, detection, and hallucination. The benchmark ultimately covers 16 DeepFake types, 3 modalities, 55K samples, and evaluates 18 MLLMs.

graph TD
    A["30+ Public Real Face Datasets<br/>Image / Video / Audio"]
    A --> B["Data Generation & QC<br/>50+ Models for 16 Types · ≥3 Models/Type · Paired Samples · Auto-QC"]
    B --> C["Fine-grained Artifact Taxonomy + Manual Annotation<br/>Quality Artifacts (Localized) / Semantic Artifacts → GT List"]
    C --> EVAL
    subgraph EVAL["Perception–Detection–Hallucination Triple-Dimension Evaluation"]
        direction TB
        D["Three Question Types<br/>TFQ / MCQ / OEQ"] --> E["18 MLLMs Answers<br/>+ External LLM Alignment to Artifact List"]
        E --> F["Perception · Fake Samples Only<br/>Acc / MCQ Scoring / Cover"]
        E --> G["Detection · Mixed OEQ-B<br/>Judge then List Artifacts · Acc / Cover"]
        E --> H["Hallucination · OEQ-A/B Fictional Artifacts<br/>CHAIR / Hal / F0.5"]
    end

Key Designs

1. Paired Real-Fake Data Generation and QC: Ensuring Accurate Annotation and Measurable Hallucination

The first step is establishing solid data. TriDF collects real faces from 30+ public datasets and generates forged samples using 50+ specialized models (GAN, SD, DiT, commercial APIs, etc.). Each forgery type is generated by at least 3 different models to ensure generator diversity, with automatic quality filtering based on authenticity and consistency metrics. The 16 DeepFake types are divided into two families: Partial Manipulation (Face Swap, Attribute Editing, Lip Sync, Face Reenactment, Full-body Manipulation, Subject-driven Editing, Voice Conversion) and Full Synthesis (Audio-driven Talking Head, Identity-preserving Generation, Text-to-Human Image/Video, etc.). Pairing real and fake samples is crucial: it allows annotators to locate artifacts by comparison, and more importantly, it supports hallucination evaluation—verifying if a model "sees" artifacts in an inherently flawless real sample.

2. Fine-grained Artifact Taxonomy: An Objective Yardstick for Perception

To solve the problem of scoring model explanations objectively, TriDF uses a two-layer human-annotated taxonomy: Quality Artifacts (blur, noise, flickering, etc., which are low-level image degradations) and Semantic Artifacts (anatomical inconsistency, object integrity defects, unnatural prosody, etc., which require common-sense reasoning). Quality artifacts are further localized to specific parts (nose, limbs, background, etc.) to evaluate the model's localization capability. Since these annotations are entirely human-generated, the perception dimension has an objective ground truth independent of MLLM self-evaluation.

3. Perception–Detection–Hallucination Evaluation: Decoupled Measurement with Specific Metrics

The framework decouples these three capabilities using different sample subsets and question types. Perception uses only fake samples via TFQ/MCQ/OEQ-A. In MCQ, "None of the above" and multi-select options are included to increase difficulty. Detection uses mixed real/fake samples via OEQ-B, requiring the model to provide a binary judgment followed by supporting artifacts, scored by Accuracy and Cover (hit rate of correct artifacts). Hallucination is quantified by identifying fictional artifacts in OEQ-A and OEQ-B responses that do not exist in the ground truth list.

To ensure objective scoring, perception and detection use Accuracy (TFQ), Cover, and a penalty-based MCQ scoring system: with \(K\) correct and \(M-K\) incorrect options, selecting a correct one yields \(+1/K\) while an incorrect one yields \(-1/(M-K)\). Hallucination is measured by CHAIR (proportion of fictional artifacts), Hal (proportion of responses with at least one hallucination), and \(F_{0.5}\). A critical penalty rule: if the model identifies zero artifacts or misclassifies a fake sample as real, CHAIR is set to 1 to prevent models from avoiding penalties by remaining silent. Alignment between free-text and the artifact list is handled by an external lightweight LLM (Gemini 1.5 Flash-Lite) to avoid self-preference bias.

Key Experimental Results

Main Results - Perception Evaluation (TFQ)

MLLM	Image TFQ Avg	Video TFQ Avg	Total Avg	Rank
GPT-5	63.36%	57.02%	60.19%	1
Gemini 2.5-Pro	61.69%	57.58%	59.63%	3
Qwen3-VL-30B	61.04%	58.65%	59.85%	2
Claude Sonnet 4.5	53.57%	51.05%	52.31%	14
InternVL3_5-8B	53.69%	54.03%	53.86%	7

Main Results - Detection + Hallucination Evaluation (Type-B OEQ)

MLLM	Image Acc	Image Cover↑	Image CHAIR↓	Image F0.5↑
Qwen3-Omni-30B	0.6942	0.4143	0.6701	0.3381
Qwen3-VL-30B	0.6894	0.3661	0.7137	0.2388
InternVL2_5-38B	0.5747	0.2306	0.8066	0.1971
GPT-4o (Proxy)	0.6521	0.3422	0.7255	0.2541

Key Findings

• Perception is the foundation of detection: Models with high perception capability (TFQ/MCQ ranks) generally show better detection performance, though it is not a sufficient condition.
• Hallucination severely undermines decision-making: Even with strong perception, high hallucination rates lead to unstable detection. Most MLLMs have a CHAIR > 0.5, meaning over half of their responses contain fabricated artifacts.
• Open-source vs. Closed-source Gap: GPT-5 ranks first in perception, but Qwen3-VL-30B performs best among open-source models. While Claude Sonnet 4.5 ranks lower in perception (14), it scores highest in MCQ (0.21), indicating higher reasoning "precision."
• Video is harder than image: Nearly all models perform worse on video modalities, indicating that temporal artifact recognition remains a challenge.
• Hallucinations are pervasive: Most models exhibit Type-A OEQ Hal > 0.9, meaning over 90% of responses contain at least one hallucinated artifact. This poses a serious threat to the credibility of interpretable detection.

Highlights & Insights

• Integrity of the Triadic Framework: Unlike previous benchmarks focusing solely on accuracy or explanation quality, TriDF integrates Perception-Detection-Hallucination, revealing their inseparable nature. This framework is transferable to other XAI fields.
• Human-annotated Artifact Taxonomy: By avoiding MLLM self-evaluation, it establishes an objective benchmark. The two-tier design of quality and semantic artifacts is highly systematic.
• Timely Introduction of Hallucination Evaluation: The hallucination problem in DeepFake detection was previously ignored, but results showing Hal > 0.9 suggest current interpretable detection is far from reliable.

Limitations & Future Work

• Artifact annotation relies on humans, which is costly and subject to inter-annotator consistency issues.
• While 55K samples is substantial, the distribution across 16 DeepFake types may be uneven, with potential data scarcity for rare types.
• The framework primarily targets static detection and does not consider interactive scenarios (e.g., follow-up questioning).
• The Cover metric only evaluates coverage, not the precision or detail of the descriptions.
• No specific suggestions or methods for mitigating hallucinations are provided.

Related Work & Insights

• vs FakeBench: FakeBench covers only 1 DeepFake type and lacks hallucination evaluation; TriDF covers 16 and includes complete hallucination assessments.
• vs LOKI: LOKI supports multi-modal but covers only 3 types and lacks a human-annotated taxonomy.
• vs Forensics-Bench: Covers 10 types but its 63K samples lack perception and hallucination evaluations.
• vs DD-VQA: Pioneered VQA-style detection but covers only 4 types; TriDF is significantly more comprehensive.

Rating

• Novelty: ⭐⭐⭐⭐ The tri-dimensional framework and hallucination evaluation are new contributions, though the core is a benchmark rather than a novel algorithm.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Covers 16 types and 51 generators, evaluating 18 MLLMs.
• Writing Quality: ⭐⭐⭐⭐ Structured clearly, though some core insights could be more prominent amidst the numerous tables.
• Value: ⭐⭐⭐⭐⭐ Fills a critical gap in evaluating interpretable DeepFake detection.

Related Papers

• [CVPR 2026] Evaluating and Easing Hallucinations for GUI Grounding
• [CVPR 2026] Lyapunov Probes for Hallucination Detection in Large Foundation Models
• [CVPR 2026] Zina: Multimodal Fine-grained Hallucination Detection and Editing
• [ICML 2026] From Out-of-Distribution Detection to Hallucination Detection: A Geometric View
• [ICLR 2026] Beyond In-Domain Detection: SpikeScore for Cross-Domain Hallucination Detection