Unified Multimodal Fake Content Detection

• UMFDet is a unified framework that detects fake content across heterogeneous modalities using a single end-to-end model.
• It leverages various architectural families, including fusion-centric and modality-agnostic designs, to address cross-modal inconsistencies and deception taxonomies.
• Benchmark results and robust training protocols highlight its strengths in classification and explanation while addressing challenges like out-of-distribution generalization and modality bias.

Unified Multimodal Fake Content Detection (UMFDet) denotes a class of methods that aim to detect deceptive content across heterogeneous modalities within a single framework rather than through separate task-specific detectors. In one explicit formulation, it is a single, end-to-end model that accepts multiple input modalities, handles heterogeneous fake types, and jointly outputs a global authenticity decision, localization, and a natural-language explanation (Li et al., 2 May 2026). A closely related formulation targets multimodal social-media posts and assigns one of three labels—Real, Human-crafted misinformation, or AI-synthesized—through a unified vision–LLM with category-specific specialization (Li et al., 30 Sep 2025). Within this umbrella, multimodal fake news detection with paired text–image inputs is a special case, and modality-agnostic audio–visual deepfake detection is another (Zhou et al., 2 Jun 2026, Yu et al., 2023).

1. Emergence of the unified formulation

UMFDet did not emerge from a single lineage. Earlier work concentrated on text–image fake news detection and progressively enriched fusion granularity, semantic alignment, and auxiliary cues. Representative examples include EM-FEND, which models entity inconsistency, mutual enhancement, and text complementation (Qi et al., 2021); ETMA, which comprises a visual attention-based encoder, textual attention-based encoder, and joint attention-based learning (Yadav et al., 2022); MMFN, which fuses fine-grained features while taking into account coarse-grained features encoded by the CLIP encoder (Zhou et al., 2023); and CFFN, which separately explores consistency and inconsistency from high-relevant and low-relevant word-region pairs (Li et al., 2023).

System	Scope	Distinctive contribution
EM-FEND (Qi et al., 2021)	Text + image fake news	entity inconsistency, mutual enhancement, text complementation
MMFN (Zhou et al., 2023)	Text + image fake news	multi-grained fusion with fine-grained features and CLIP coarse-grained features
CFFN (Li et al., 2023)	Text + image fake news	consistency and inconsistency from high-relevant and low-relevant word-region pairs
Modality-agnostic deepfake detector (Yu et al., 2023)	audio, video, audio+video	dual-label detection and missing-modality handling
UMFDet (Li et al., 30 Sep 2025)	social-media image–text deception	Category-aware Mixture-of-Experts Adapter and attribution chain-of-thought
Omni-Fake-R1 (Li et al., 2 May 2026)	image, audio, video, audio–video talking head	joint detection–localization–explanation

A second strand arose from multimodal deepfake detection, where the emphasis shifted from text–image consistency to modality-agnostic handling of audio and video. The framework in "A Unified Framework for Modality-Agnostic Deepfakes Detection" explicitly targets audio-only, video-only, and paired audio–video inputs, and predicts per-modality authenticity labels rather than a single pair-level verdict (Yu et al., 2023). More recent work extends unification further: DeepfakeBench-MM standardizes multimodal audiovisual benchmarking (Zhao et al., 26 Oct 2025), Omni-Fake defines a unified omni-dataset and a joint detection–localization–explanation protocol across four modalities (Li et al., 2 May 2026), and OmniFake with UMFDet explicitly bridges human-crafted misinformation and AI-generated multimodal content in one ternary framework (Li et al., 30 Sep 2025).

2. Task formulations and deception taxonomies

The narrowest formulation within this area is multimodal fake news detection (MFND): given a paired text–image news item $(T, I)$, predict whether it is fake or real. IDO treats this setting on Weibo, Weibo-21, and GossipCop, and explicitly describes MFND as a special case within the broader concept of UMFDet (Zhou et al., 2 Jun 2026). Earlier text–image work on Fakeddit further shows that the label space need not be binary: one six-class taxonomy comprises True, Manipulated Content, False Connection, Satire/Parody, Misleading Content, and Imposter Content (Alonso-Bartolome et al., 2021).

Audiovisual deepfake work broadens the label structure in a different way. DeepfakeBench-MM defines Multimodal DeepFake Detection on audiovisual clips $(x^a, x^v)$ and further distinguishes RARV, RAFV, FARV, and FAFV, making modality-specific authenticity explicit rather than collapsing all manipulation into a single fake label (Zhao et al., 26 Oct 2025). The modality-agnostic framework of (Yu et al., 2023) turns this into dual-label detection with $y_a \in \{0,1\}$, $y_v \in \{0,1\}$, and a fake-composition label $p = y_a + y_v$, thereby encoding not only whether content is fake but also which modality is fake.

The most expansive formulations add both heterogeneous modalities and heterogeneous outputs. Omni-Fake defines a unified task over Image, Audio, Generic video, and Audio–video talking head inputs. Depending on modality, the global label is either $$\{\text{REAL}, \text{TAMPERED}, \text{FULL\_SYNTHETIC}\}$$ or $\{\text{REAL}, \text{FAKE}\}$, and the model is further expected to emit spatial or temporal localization and a natural-language rationale in a structured format (Li et al., 2 May 2026). OmniFake and UMFDet adopt a different but related ternary taxonomy—Real, Human-crafted misinformation, and AI-synthesized—motivated by the observation that real-world social-media posts do not announce whether deception is human-authored or generator-produced (Li et al., 30 Sep 2025).

This plurality of task definitions shows that “unified” has at least three meanings in the literature: unification across modalities, unification across deception types, and unification across outputs such as classification, localization, and explanation.

3. Architectural families

One major architectural family remains fusion-centric. EM-FEND extracts visual entities, embedded text in images, and visual CNN features, then models entity inconsistency, mutual enhancement, and text complementation before late representation-level fusion (Qi et al., 2021). ETMA uses a visual attention-based encoder, a textual attention-based encoder, and a joint attention-based learning module, with visual semantic attention and self-attention over fused features (Yadav et al., 2022). MMFN explicitly targets multi-grained information fusion by combining token-level features from Transformer-based text and image encoders with coarse-grained CLIP features and similarity-based weighting for ambiguity handling (Zhou et al., 2023). CFFN divides word–region pairs into high-relevant and low-relevant parts, learns consistency from the former and inconsistency from the latter, and uses a selection module to choose the primary clue (Li et al., 2023).

A second family focuses on why multimodal fusion fails. IDO argues that cross-modal consistency alone is insufficient and introduces two distinct notions: factual incongruity and modality incongruity. Its architecture combines ViT and BERT backbones, cross-modal attention, a Factual Semantic Distribution module with Gaussian distribution modeling over memory banks, and Incongruity Contrastive Learning. The final decision is $$\hat{y} = \sigma(A_{\text{FSD} } + X_{\text{ICL} })$$, directly combining factual and modality incongruity scores (Zhou et al., 2 Jun 2026). FND-MoE introduces “modality disruption” to describe disruptive or over-expressive information in certain modalities and addresses it through modality-specific experts, pairwise cross-attention, and a two-pass feature selection mechanism based on Top-k selection and Gumbel-Sigmoid gating (Liu et al., 12 Apr 2025). CIMDD instead frames multimodal fake news detection causally: it posits Lexical Semantic Confounder, Latent Visual Confounder, and Dynamic Cross-Modal Coupling Confounder, then counters them through backdoor adjustment, frontdoor adjustment, and cross-modal joint intervention (Liu et al., 12 Apr 2025).

A third family redefines unification itself. The modality-agnostic audio–visual framework of (Yu et al., 2023) uses audio–visual speech recognition as pretraining, a Modality Compensation Adapter, and a Dual-Label Classifier so that the same model handles audio-only, video-only, and audio–video inputs. UNITE-FND pursues a radically different path: it reframes multimodal fake news detection as a unimodal text classification task by using Gemini 1.5 Pro to convert images into structured textual descriptions, then trains text-only classifiers on the merged sequence $T_{final} = f_{merge}(T_{clean}, T_{desc})$ (Mukherjee et al., 16 Feb 2025). DIFND combines conditional latent diffusion over debunk textual embeddings with a chain-of-debunk multi-agent MLLM system, then fuses original text, diffusion-generated debunk features, audio, visual, and reasoning features inside a unified architecture (Yan et al., 11 Jun 2025). UMFDet proper uses a Florence-2 backbone augmented with a Category-aware Mixture-of-Experts Adapter and an Attribution Chain-of-Thought mechanism (Li et al., 30 Sep 2025), whereas Omni-Fake-R1 is built atop Qwen2.5-Omni-7B and uses curriculum SFT with modal replay followed by GSPO reinforcement learning for structured detection, localization, and explanation (Li et al., 2 May 2026).

Taken together, these systems show that unification is not tied to a single implementation pattern. Some models unify by adding more modalities to a common backbone; some unify by imposing common reasoning structures; some unify by translating all modalities into text.

4. Benchmarks and empirical landscape

Benchmark construction has become central because unification claims are only meaningful when evaluated across heterogeneous manipulations and out-of-distribution conditions.

Benchmark	Modalities	Notable property
Fakeddit (Alonso-Bartolome et al., 2021)	text + image	over 1 million instances; six-class taxonomy
Weibo / Weibo-21 / GossipCop (Zhou et al., 2 Jun 2026)	text + image	standard MFND benchmarks
FakeSV / FVC (Liu et al., 12 Apr 2025)	text, image, audio, video	short-video and news-style multimodal fake news
Mega-MMDF / DeepfakeBench-MM (Zhao et al., 26 Oct 2025)	audio + video	1.2M clips; 21 forgery pipelines; standardized benchmark
Omni-Fake (Li et al., 2 May 2026)	image, audio, video, AV-TH	Omni-Fake-Set ≈ 1.137M and Omni-Fake-OOD ≈ 209k; detection–localization–explanation
OmniFake (Li et al., 30 Sep 2025)	image + text	127,283 ternary samples integrating human-crafted and AI-generated deception

Results are correspondingly heterogeneous. On Fakeddit, the multimodal CNN in "Multimodal Fake News Detection" reports 87% accuracy, while the best unimodal BERT baseline reports 78%, and Manipulated Content, False Connection, and Satire/Parody benefit strongly from images (Alonso-Bartolome et al., 2021). On standard text–image MFND benchmarks, IDO reports Accuracy 0.947 on Weibo, 0.912 on GossipCop, and 0.963 on Weibo-21 (Zhou et al., 2 Jun 2026). On OmniFake, UMFDet reports 86.67 ACC and 86.73 F1 overall, with 95.39 F1 on AI-synthesized samples (Li et al., 30 Sep 2025). On Omni-Fake-OOD, Omni-Fake-R1 reports 79.25 image accuracy, 85.96 video accuracy, 83.85 audio accuracy, and 86.52 AV-TH accuracy, while also outputting localization and explanation (Li et al., 2 May 2026).

The benchmark picture also complicates simplistic narratives about foundation models. DeepfakeBench-MM reports that zero-shot Qwen2.5-Omni and VideoLLaMA2 achieve 0.568 and 0.534 AUC on Mega-MMDF, indicating that generic MLLMs are not competitive on multimodal deepfake detection without task-specific adaptation (Zhao et al., 26 Oct 2025). By contrast, UNITE-FND shows that if multimodal inputs are first translated into structured text, a text-only pipeline can be very strong: on Uni-Fakeddit-55k, DeBERTa-v3-large with Structured Image Description reaches 92.52% binary accuracy while using a downstream classifier rather than direct multimodal fusion (Mukherjee et al., 16 Feb 2025).

5. Optimization objectives and evaluation protocols

Cross-entropy remains the dominant base objective, but unified systems increasingly add structure-specific losses. Classic text–image classifiers and ETMA use standard softmax or binary cross-entropy classification losses (Alonso-Bartolome et al., 2021, Yadav et al., 2022). CFFN supplements detection loss with a partition loss, $\mathcal{L} = \mathcal{L}_d + \beta \mathcal{L}_p$, where $(x^a, x^v)$0 for real news and $(x^a, x^v)$1 for fake news, thereby supervising the selector to rely on consistency for real and inconsistency for fake; the reported settings are $(x^a, x^v)$2 on Twitter and $(x^a, x^v)$3 on Weibo (Li et al., 2023). IDO uses $(x^a, x^v)$4 with $(x^a, x^v)$5, combining supervised classification with continuous contrastive learning (Zhou et al., 2 Jun 2026). DIFND further extends the objective to $(x^a, x^v)$6, explicitly optimizing multimodal fusion, textual-debunk fusion, and diffusion-based evidence generation together (Yan et al., 11 Jun 2025). UMFDet couples answer supervision and reasoning supervision through $(x^a, x^v)$7 (Li et al., 30 Sep 2025).

The most elaborate training protocol in this group is Omni-Fake-R1. It first uses curriculum SFT with modal replay in the order Audio, then Audio+Image, then Audio+Image+Video, then Audio+Image+Video+AV-TH, retaining 15% replay from earlier modalities at each stage. It then applies GSPO reinforcement learning with a unified scalar reward,

$(x^a, x^v)$8

with $(x^a, x^v)$9, $y_a \in \{0,1\}$0, $y_a \in \{0,1\}$1, and $y_a \in \{0,1\}$2, thereby aligning a single policy with structured output format, classification correctness, spatial localization, and temporal localization (Li et al., 2 May 2026).

Evaluation protocols have expanded in parallel. Standard MFND papers report Accuracy and class-wise F1; IDO reports Accuracy together with F1(Fake) and F1(Real) (Zhou et al., 2 Jun 2026). DeepfakeBench-MM emphasizes AUC, ACC, AP, and EER, together with intra-dataset, cross-dataset, and cross-pipeline evaluation (Zhao et al., 26 Oct 2025). Omni-Fake adds IoU and localization F1 for images, videos, and audio intervals, while explanation quality is measured by ROUGE-L, CSS, and expert 1–5 Likert ratings (Li et al., 2 May 2026). UNITE-FND additionally introduces IPR, SCS, ISS, SIR, MTE, and CIQS to evaluate the quality of multimodal-to-unimodal translation (Mukherjee et al., 16 Feb 2025). This diversification of metrics reflects the field’s shift from plain classification accuracy toward explanation, localization, and robustness.

6. Robustness, interpretability, and unresolved issues

A persistent misconception in the area is that adding modalities is necessarily beneficial. "Is Multi-Modal Necessarily Better?" directly challenges this view: under five adversarial and two backdoor attack methods, the performance of state-of-the-art multimodal detectors degrades significantly, most detectors are more vulnerable to attacks on the visual modality than the textual modality, popular events’ images cause significant degradation under backdoor attacks, performance under multimodal attacks is worse than under unimodal attacks, and defensive methods improve robustness (Chen et al., 2022). A related but milder claim appears in FND-MoE: adding all available modalities can be worse than using only a subset, because certain modalities contain disruptive or over-expressive information (Liu et al., 12 Apr 2025).

Modality bias is also repeatedly documented. DeepfakeBench-MM shows that detectors perform much better on video-only than audio-only input and identifies strong visual dominance; random modality masking improves balance and cross-dataset AUC for AVTS and MRDF (Zhao et al., 26 Oct 2025). The modality-agnostic framework of (Yu et al., 2023) tackles the same issue with a Modality Compensation Adapter and dual-label supervision, allowing the detector to remain usable when one modality is absent. This suggests that robustness in UMFDet is not only an attack-resistance problem but also a representation-balancing problem.

Interpretability has become a second major axis of unification. CFFN’s selector explicitly chooses whether consistency or inconsistency is the primary clue (Li et al., 2023). DIFND generates debunking evidence and reasoning text through diffusion and multi-agent MLLM inference (Yan et al., 11 Jun 2025). UMFDet trains on attribution-style chain-of-thought targets (Li et al., 30 Sep 2025). Omni-Fake-R1 makes explanation a first-class output rather than a post hoc add-on (Li et al., 2 May 2026). At the same time, current systems still face clear limitations: IDO depends on labeled fake/real instances, remains two-modal, uses internal-only factual modeling, and assumes single Gaussians for class semantics (Zhou et al., 2 Jun 2026); Omni-Fake currently does not cover 3D avatars or text modality (Li et al., 2 May 2026); OmniFake is derived primarily from Reddit and English-language content and collapses many AI manipulation types into one category (Li et al., 30 Sep 2025).

Three unresolved issues therefore dominate the present research frontier. First, out-of-distribution generalization remains difficult even on newly standardized benchmarks. Second, zero-shot generic MLLMs are not yet sufficient, despite their broad multimodal competence (Zhao et al., 26 Oct 2025). Third, there is no consensus on whether the most scalable path to unification is a large native multimodal backbone, a modality-agnostic representation pipeline, or a translation-based substrate such as UNITE-FND (Mukherjee et al., 16 Feb 2025). This suggests that UMFDet is best understood not as a settled architecture but as an evolving systems objective: one model, heterogeneous modalities, heterogeneous deception types, and increasingly structured outputs under realistic robustness constraints.