Harmful Visual Content Manipulation Matters in Misinformation Detection Under Multimedia Scenarios

Bing Wang; Ximing Li; Changchun Li; Jinjin Chi; Tianze Li; Renchu Guan; Shengsheng Wang

cs.LGMar 22, 2026

Harmful Visual Content Manipulation Matters in Misinformation Detection Under Multimedia Scenarios

Authors: Bing Wang, Ximing Li, Changchun Li, Jinjin Chi, Tianze Li, Renchu Guan, Shengsheng Wang

Organizations: College of Computer Science and Technology, Jilin University, China · School of Advanced Technology, Xi’an Jiaotong-Liverpool University, China

Abstract

Nowadays, the widespread dissemination of misinformation across numerous social media platforms has led to severe negative effects on society. To address this challenge, the automatic detection of misinformation, particularly under multimedia scenarios, has gained significant attention from both academic and industrial communities, leading to the emergence of a research task known as Multimodal Misinformation Detection (MMD). Typically, current MMD approaches focus on capturing the semantic relationships and inconsistency between various modalities but often overlook certain critical indicators within multimodal content. Recent research has shown that manipulated features within visual content in social media articles serve as valuable clues for MMD. Meanwhile, we argue that the potential intentions behind the manipulation, e.g., harmful and harmless, also matter in MMD. Therefore, in this study, we aim to identify such multimodal misinformation by capturing two types of features: manipulation features, which represent if visual content has been manipulated, and intention features, which assess the nature of these manipulations, distinguishing between harmful and harmless intentions. Unfortunately, the manipulation and intention labels that supervise these features to be discriminative are unknown. To address this, we introduce two weakly supervised indicators as substitutes by incorporating supplementary datasets focused on image manipulation detection and framing two different classification tasks as positive and unlabeled learning issues. With this framework, we introduce an innovative MMD approach, titled Harmful Visual Content Manipulation Matters in MMD (HAVC-M4 D). Comprehensive experiments conducted on four prevalent MMD datasets indicate that HAVC-M4 D significantly and consistently enhances the performance of existing MMD methods.

Figures & tables

Fig. 1: The statistics from the MMD dataset Twitter reveal a quantitative relationship between manipulated visual content and veracity labels. We utilize a pre-trained model for image manipulation detection to determine if an image has been manipulated. Additionally, we provide various examples of images that have been manipulated with both harmful and harmless intentions.

Notation	Description
$\mathbf{x}_{i}^{t}$	text content
$\mathcal{X}_{i}^{v}=\{\mathbf{x}_{ij}^{v}\}_{j=1}^{K}$	visual content
$y_{i}$	veracity label
$N$	number of samples
$K$	number of images
$\boldsymbol{\theta}^{t},\boldsymbol{\theta}^{v}$	parameters of text and visual encoders

TABLE I: Summary of notations.

Fig. 2: The Havc-m 4 d framework comprises four primary encoders: text encoder, image encoder, manipulation encoder, and intention encoder, which extract features from the provided text content $\mathbf{x}_{i}^{t}$ and visual content $\mathcal{X}_{i}^{v}$ . These features are subsequently fed into a feature fusion network to create a combined representation. To accomplish the multiple tasks, we employ three distinct predictors aimed at veracity classification, manipulation classification, and intention classification.

Dataset	# Real	# Fake	# Images
GossipCop [ 22 ]	10,259	2,581	12,840
Weibo [ 12 ]	4,779	4,749	9,528
Twitter [ 23 ]	6,026	7,898	514
FakeSV [ 24 ]	1,827	1,827	3,654

TABLE II: Statistics of three prevalent MMD datasets.

Method	Accuracy	Macro F1	Real			Fake			Avg. $\Delta$
			Precision	Recall	F1	Precision	Recall	F1
Dataset : GossipCop
Base model	87.77 $\pm$ 0.56	79.51 $\pm$ 0.44	91.55 $\pm$ 0.41	93.36 $\pm$ 1.20	92.37 $\pm$ 0.41	69.96 $\pm$ 1.10	63.30 $\pm$ 1.46	66.92 $\pm$ 0.58	-
+ Havc-m 4 d	88.45 $\pm$ 0.20 *	80.32 $\pm$ 0.43 *	91.93 $\pm$ 0.14	94.08 $\pm$ 0.33 *	92.83 $\pm$ 0.12	71.99 $\pm$ 0.80 *	64.59 $\pm$ 0.74 *	67.63 $\pm$ 0.78 *	+0.90
SAFE [ 64 ]	87.78 $\pm$ 0.31	79.22 $\pm$ 0.49	91.22 $\pm$ 0.30	93.34 $\pm$ 0.47	92.37 $\pm$ 0.20	70.66 $\pm$ 1.32	63.12 $\pm$ 1.50	66.66 $\pm$ 0.84	-
+ Havc-m 4 d	88.53 $\pm$ 0.24 *	79.87 $\pm$ 0.30 *	91.90 $\pm$ 0.31 *	94.32 $\pm$ 0.54 *	92.95 $\pm$ 0.20 *	72.19 $\pm$ 1.30 *	64.44 $\pm$ 0.73 *	67.88 $\pm$ 0.51 *	+0.96

TABLE III: Experimental results of Havc-m 4 d across three prevalent MD datasets GossipCop , Weibo , and Twitter . The results marked by * are statistically significant compared to its baseline models, satisfying p-value

<

0.05.

Method	Accuracy	Macro F1	Precision	Recall	AUC	Avg. $\Delta$
VGG [ 65 ]	69.45 $\pm$ 2.12	64.52 $\pm$ 1.96	68.08 $\pm$ 2.87	64.19 $\pm$ 1.84	64.19 $\pm$ 1.84	-
+ Havc-m 4 d	70.85 $\pm$ 1.31 *	66.12 $\pm$ 2.35 *	70.00 $\pm$ 0.78 *	65.82 $\pm$ 1.91 *	65.82 $\pm$ 1.91 *	+1.63
C3D [ 66 ]	67.60 $\pm$ 1.62	67.43 $\pm$ 1.59	67.97 $\pm$ 1.67	67.60 $\pm$ 1.58	67.60 $\pm$ 1.58	-
+ Havc-m 4 d	68.80 $\pm$ 1.77 *	68.70 $\pm$ 1.79 *	68.98 $\pm$ 1.77 *	68.77 $\pm$ 1.78 *	68.77 $\pm$ 1.78 *	+1.16
FANVM [ 67 ]	77.14 $\pm$ 3.07	77.11 $\pm$ 3.05	77.26 $\pm$ 3.14	77.13 $\pm$ 3.04	77.13 $\pm$ 3.04	-
+ Havc-m 4 d	78.39 $\pm$ 2.44 *	78.38 $\pm$ 2.42 *	78.48 $\pm$ 2.54 *	78.39 $\pm$ 2.43 *	78.39 $\pm$ 2.43 *	+1.25

TABLE IV: Experimental results of Havc-m 4 d across the prevalent MMD dataset FakeSV . The results marked by * are statistically significant compared to its baseline models, satisfying p-value

<

0.05.

Method	Dataset : GossipCop					Dataset : Weibo					Dataset : FakeSV
	Acc.	F1	AUC	F1 ${}_{\text{real}}$	F1 ${}_{\text{fake}}$	Acc.	F1	AUC	F1 ${}_{\text{real}}$	F1 ${}_{\text{fake}}$	Acc.	F1	P.	R.	AUC
SOTA + Havc-m 4 d	87.95	79.99	87.46	93.14	66.80	91.74	91.68	97.06	91.56	91.81	80.67	80.63	80.91	80.66	80.66
w/o $\mathcal{L}_{PRE}$	87.24	79.18	87.21	92.14	66.23	90.17	90.12	96.93	89.91	90.83	79.45	79.41	79.70	79.47	79.47
w/o $\mathcal{L}_{PU}$	87.63	79.66	87.38	93.05	66.52	91.40	91.39	96.96	91.19	91.60	80.35	80.32	80.52	80.36	80.36
w/o $\mathcal{L}_{KD},\mathcal{L}_{IR}$	86.93	78.65	86.24	91.94	65.36	90.10	90.10	95.97	89.88	90.31	78.98	78.94	79.21	79.00	79.00

TABLE V: Ablative study on objective functions. The bold and underlined scores indicate the highest and lowest results in the ablative versions, respectively.

Method	Dataset : GossipCop					Dataset : Weibo					Dataset : FakeSV
	Acc.	F1	AUC	F1 ${}_{\text{real}}$	F1 ${}_{\text{fake}}$	Acc.	F1	AUC	F1 ${}_{\text{real}}$	F1 ${}_{\text{fake}}$	Acc.	F1	P.	R.	AUC
SOTA + Havc-m 4 d	87.95	79.99	87.46	93.14	66.80	91.74	91.68	97.06	91.56	91.81	80.67	80.63	80.91	80.66	80.66
w/o $\mathbf{e}^{m}$	87.53	79.13	86.47	92.37	65.89	90.92	90.91	96.52	90.57	91.08	79.77	79.72	79.93	79.79	79.79
w/o $\mathbf{e}^{e}$	87.69	79.56	86.72	92.39	66.25	91.13	91.11	96.78	90.78	91.45	80.29	80.23	80.40	80.26	80.26
w/o $\mathbf{e}^{m},\mathbf{e}^{e}$	87.26	79.03	86.27	92.14	65.51	90.17	90.15	96.45	89.81	90.50	79.03	78.99	79.94	79.17	79.17

TABLE VI: Ablative study on manipulation and intention features.

Fig. 3: Sensitivity analysis of the parameters

\alpha

and

\beta

Fig. 4: Visualization analysis of features

\mathbf{z}

\mathbf{e}^{m}

and

\mathbf{e}^{e}

with the T-SNE method.

Fig. 5: We illustrate three representative examples for the case study.

Methods	Base	SAFE	MCAN	CAFE	BMR
GossipCop	20.2	23.0	28.8	18.8	23.8
+ Havc-m 4 d	27.4	27.8	34.6	23.6	27.8
$\Delta$	1.35 $\times$	1.20 $\times$	1.20 $\times$	1.25 $\times$	1.16 $\times$
Weibo	8.0	10.0	15.8	10.0	10.6
+ Havc-m 4 d	12.0	14.2	18.8	13.8	14.0
$\Delta$	1.50 $\times$	1.42 $\times$	1.18 $\times$	1.38 $\times$	1.32 $\times$

TABLE VII: Training budget of our proposed Havc-m 4 d .

Figure 13

Figure 14

Figure 15

Figure 16

Figure 17

Figure 18

Fake News Detection

Fake news detection identifies false or misleading claims by analyzing content, context, evidence, and how information spreads.

See all 98 papers

What Improves Multimodal Misinformation Detection? Answers from a Large-Scale Empirical Study
Sep 24, 2026Akshit Sharma, Prashant W. PatilMultimodal Fusion Multimodal Model Evaluation
CVPR Lab, MFSDSAI Indian Institute of Technology Guwahati
Probabilistic Concept Graph Reasoning for Multimodal Misinformation Detection
Mar 26, 2026Ruichao Yang, Wei Gao, Xiaobin Zhu +5Graph Reasoning Multimodal Robustness
University of Science and Technology Beijing · Singapore Management University · Hong Kong Baptist University
ReMMD: Realistic Multilingual Multi-Image Agentic Verification for Multimodal Misinformation Detection
Jun 23, 2026Chenhao Dang, Dantong Zhu, Jun Yang +2Multimodal Agents Multimodal Model Evaluation
Shanghai Jiaotong University · Shanghai Artificial Intelligence Laboratory · Central South University +2

Sep 24, 2026J/K move · Enter open · S save

What Improves Multimodal Misinformation Detection? Answers from a Large-Scale Empirical Study

Akshit Sharma, Prashant W. Patil

CVPR Lab, MFSDSAI Indian Institute of Technology Guwahati

Multimodal Fusion Multimodal Model Evaluation Multimodal Robustness

Open paper AskPDF arXiv

Multimodal misinformation is increasingly crafted to look convincing by pairing a textual claim with an image that appears to "prove" it. Yet in practice, building effective detectors often hinges on a small set of design choices that are rarely examined in a controlled way. In this paper, we conduct a large-scale study of multimodal design choices for misinformation detection with over 3,375 experiments- spanning three benchmark datasets and a broad range of pre-trained vision and language backbones. Through systematic comparisons and targeted robustness analyses, we distill practical guidance on which design choices help, when do they fail silently, and what aspects of the pipeline most strongly shape model behavior, answering 4 key Research Questions (RQs). We aim to provide a reliable foundation for designing stronger and more dependable multimodal misinformation detection systems, thus contributing to the broader research community.

Abstract

Fake News Detection

Explore similar work

What Improves Multimodal Misinformation Detection? Answers from a Large-Scale Empirical Study

Probabilistic Concept Graph Reasoning for Multimodal Misinformation Detection

ReMMD: Realistic Multilingual Multi-Image Agentic Verification for Multimodal Misinformation Detection