PaperlayerView exact revision on arXiv

Source-linked AI summary

Exploiting Multi-domain Visual Information for Fake News Detection

Peng Qi, Juan Cao, Tianyun Yang, Junbo Guo, Jintao Li

arXiv:1908.04472v1cs.MMcs.IRcs.SI

TL;DR

Fake-news images may be either manipulated or misleadingly reused, creating a need to model their physical and semantic characteristics. MVNN combines frequency- and pixel-domain visual representations with attention, achieving higher accuracy than existing methods and improving multimodal detection.

  • Problem

    Fake-news images include tampered images and real images wrongly used to represent irrelevant events, making their visual characteristics challenging to exploit for detection.

  • Method

    MVNN jointly learns frequency-domain physical features and pixel-domain semantic features, then dynamically fuses them with attention.

  • Results

    84.6% accuracy, outperforming existing approaches by at least 9.2%; MVNN also exceeds multimodal fusion methods by over 5.2% in accuracy.

  • Takeaways & Limitations

    The frequency and pixel domains are complementary, with the pixel domain playing a major role and the frequency domain providing auxiliary information.

  • Takeaways & Limitations

    Evaluation is limited to Weibo data because the existing Twitter multimedia dataset contains too few distinctive images.

Abstract

from arXiv · show

The increasing popularity of social media promotes the proliferation of fake news. With the development of multimedia technology, fake news attempts to utilize multimedia contents with images or videos to attract and mislead readers for rapid dissemination, which makes visual contents an important part of fake news. Fake-news images, images attached in fake news posts,include not only fake images which are maliciously tampered but also real images which are wrongly used to represent irrelevant events. Hence, how to fully exploit the inherent characteristics of fake-news images is an important but challenging problem for fake news detection. In the real world, fake-news images may have significantly different characteristics from real-news images at both physical and semantic levels, which can be clearly reflected in the frequency and pixel domain, respectively. Therefore, we propose a novel framework Multi-domain Visual Neural Network (MVNN) to fuse the visual information of frequency and pixel domains for detecting fake news. Specifically, we design a CNN-based network to automatically capture the complex patterns of fake-news images in the frequency domain; and utilize a multi-branch CNN-RNN model to extract visual features from different semantic levels in the pixel domain. An attention mechanism is utilized to fuse the feature representations of frequency and pixel domains dynamically. Extensive experiments conducted on a real-world dataset demonstrate that MVNN outperforms existing methods with at least 9.2% in accuracy, and can help improve the performance of multimodal fake news detection by over 5.2%.

I. INTRODUCTION

Fake-news images can mislead through tampering or incorrect contextual reuse, with distinctive physical and semantic characteristics that MVNN combines for detection.

  • Motivation: Fake-news images include digitally tampered images and authentic images wrongly used to represent unrelated events.Examples include splicing, artworks reposted as photographs, outdated crash imagery, and behavior taken out of context.
  • Motivation: Existing visual approaches struggle because handcrafted forensics target specific manipulation traces, while generic pretrained representations lack task-relevant semantic information.These limitations make fully exploiting fake-news-image characteristics important but challenging.
  • Motivation: Fake-news images may differ from real-news images at physical and semantic levels.Re-compression can produce frequency-domain artifacts, while fake-news imagery can be more visually striking and emotionally provocative.
  • Proposed framework: MVNN uses a CNN-based frequency-domain network to capture physical characteristics and a multi-branch CNN-RNN pixel-domain network to extract semantic-level features.The frequency branch transforms pixel-domain input into frequency-domain representations.
  • Proposed framework: An attention-based fusion sub-network dynamically combines frequency- and pixel-domain features because visual features contribute unequally to fake-news detection.The framework classifies the input image as fake-news or real-news imagery.

II. RELATED WORK

Prior fake-news detection methods use textual, social, statistical, forensic, or generic visual information, but these approaches have limited visual expressiveness or applicability.

  • Existing approaches: Most existing fake-news detection approaches focus on text content and social context, while visual information has only more recently become an important indicator.The growing use of multimedia has motivated incorporating images into detection systems.
  • Existing approaches: Basic image statistics such as image count, popularity, and type cannot adequately represent complex visual-content distributions.These features provide only limited descriptions of attached images.
  • Existing approaches: Handcrafted visual forensics features assist manipulation detection but are designed around specific traces and do not generalize well to misleading images.Benchmark tasks have provided multiple image-forensics feature types for manipulated and misleading multimedia use.
  • Existing approaches: Pretrained CNN representations provide general visual features but can miss semantic commonalities specific to fake-news images.Prior multimedia systems commonly fuse deep CNN features with textual information.
  • Proposed direction: MVNN addresses these limitations by modeling visual content at both physical and semantic levels for fake-news detection.Its goal is to overcome the restricted coverage of earlier visual approaches.

III. PROBLEM FORMULATION

The paper defines fake news and fake-news images for microblogging, then formulates image-based classification and introduces MVNN as the proposed solution framework.

  • III. PROBLEM FORMULATION: In microblogs, fake news is a news post intentionally and verifiably false.This differs from the traditional article-centered definition because microblog content consists of short social-media posts.
  • III. PROBLEM FORMULATION: A fake-news image is an image attached in fake news.The definition concerns the image associated with a false news post.
  • III. PROBLEM FORMULATION: The studied task uses visual content to identify a news post as real or fake, equivalently classifying its image as fake-news or real-news imagery.The formulation focuses on using the corresponding image for classification.
  • III. PROBLEM FORMULATION: Given news posts, corresponding images, and labels, the classifier learns whether a post is fake news or real news.The labels assign fake news as yt = 1 and real news as yt = 0.
  • IV. METHODOLOGY: The methodology section presents MVNN through an overview followed by detailed descriptions of its three major components.The components are the frequency-domain, pixel-domain, and fusion sub-networks.
  • IV. METHODOLOGY: MVNN feeds an input image into frequency- and pixel-domain sub-networks to obtain physical- and semantic-level features.The fusion sub-network then uses these learned representations to support classification.

B. Frequency Domain Sub-network

The frequency-domain sub-network captures physical image characteristics, while the pixel-domain sub-network extracts semantic features at multiple levels using CNN and recurrent modeling.

  • Frequency-domain processing: DCT transforms input images into frequency-domain representations to capture physical characteristics associated with tampering and re-compression.The sub-network uses block DCT before CNN-based feature extraction.
  • Frequency-domain processing: A CNN-based network automatically captures fake-news image characteristics in the frequency domain.Its architecture is shown in Fig. 5.
  • Pixel-domain processing: The pixel-domain sub-network uses multiple CNN branches to extract visual features spanning low-level to high-level semantic representations.The branches characterize image elements such as lines, colors, textures, and objects.
  • Pixel-domain processing: A bidirectional GRU models dependencies among feature levels in both local-to-global and global-to-local directions.The branch outputs are represented as a sequence and processed in forward and backward directions.

D. Fusion Sub-network

The fusion sub-network combines physical and semantic image features with attention, producing a representation for classifying images as fake-news or real-news images.

  • Feature fusion: MVNN assumes physical and semantic image features are complementary and fuses them for fake-news image detection.The fusion input includes the frequency-domain feature and pixel-domain features from multiple semantic levels.
  • Attention mechanism: An attention mechanism assigns normalized weights to feature vectors and computes an enhanced representation as their weighted sum.The feature significance scores are converted into weights through softmax.
  • Prediction: The fused representation is projected by a fully connected layer with softmax activation into fake-news and real-news classes.The output is a probability distribution over the two classes.
  • Training objective: MVNN defines training loss as cross-entropy between predicted probabilities and ground-truth labels.The label is 1 for a fake-news image and 0 for a real-news image.

V. EXPERIMENTS

The experiments evaluate MVNN on visual fake-news detection, component effectiveness, and its contribution to multimodal fake-news detection using a real-world dataset.

  • Evaluation goals: The experiments use a real-world dataset to evaluate the effectiveness of MVNN.The experimental section introduces evaluation questions before describing datasets, baselines, implementation, and comparisons.
  • Evaluation goals: EQ1 tests whether MVNN improves fake-news detection based on visual modality.This question is addressed through comparison with representative baseline methods.
  • Evaluation goals: EQ2 examines the effects of domains, attention, Bi-GRU, and pixel-domain branches through ablation studies.The analysis includes quantitative and qualitative evaluations.
  • Evaluation goals: EQ3 tests whether MVNN improves multimodal fake-news detection.The study compares multimodal detection performance after incorporating MVNN.

A. Dataset

The study evaluates MVNN on the Weibo dataset because the Twitter dataset contains too few distinctive images to support training the proposed model.

  • Dataset selection: The experiments use the Weibo dataset alone because duplicate images leave fewer than 500 distinctive images in the Twitter dataset.The authors state that Twitter is too small to support training MVNN.
  • Dataset composition: The Weibo dataset contains 4749 fake-news posts and 4779 real-news posts with corresponding images.Text-only posts are removed and posts with multiple illustrations retain one image.
  • Dataset splitting: News posts are clustered into 200 groups before splitting into training, validation, and testing sets to prevent event overlap.The split is designed to reduce overfitting on events.
  • Baselines: Baseline comparisons include forensic features with logistic regression, pre-trained VGG, fine-tuned VGG, and ConvAE.These methods represent handcrafted forensic, pretrained visual, task-finetuned, and autoencoder-based approaches.

C. Implementation Details

The experiments compare MVNN with visual baselines and evaluate how its domains and architectural components affect fake-news image detection.

  • The study evaluates fake-news image detection using accuracy, precision, recall, and F1 score for the fake-news image class.
  • Performance Comparison: 84.6% accuracy: MVNN outperforms existing approaches by at least 9.2% in single-visual-modality detection.
  • Performance Comparison: Fine-tuned VGG performs better than Pre-trained VGG, while ConvAE is slightly better than Pre-trained VGG.
  • Ablation Study: The ablation study removes the frequency domain, pixel domain, attention, Bi-GRU, branches, or combinations of these components to assess their contributions.
  • Ablation Study: The compared internal models are simplified MVNN variants, including versions without frequency-domain processing, pixel-domain processing, attention, Bi-GRU, or branches.

1) Multiple domains:

MVNN combines frequency- and pixel-domain representations because both contribute to fake-news image detection, with the pixel domain contributing more strongly and the fused representation separating classes best.

  • Multiple domains: 5.2% and 10.9%: removing the frequency- and pixel-domain sub-networks lowers accuracy by these amounts, respectively.
  • Multiple domains: The pixel domain plays a major role in detection, while the frequency domain provides auxiliary information.
  • Network Components: Attention, Bi-GRU, and pixel-domain branches all contribute to MVNN performance; removing them reduces accuracy by 1.9%, 1.8%, and 4.3%, respectively.
  • Multiple domains: MVNN, the pixel-domain sub-network, and the frequency-domain sub-network rank in that order by feature-representation separability.
  • Multiple domains: Frequency-domain representations overlap substantially, whereas MVNN produces a relatively visible boundary between fake-news and real-news samples.

F. Application on Multimodal Fake News Detection

The paper evaluates MVNN visual representations within multimodal fake-news detection systems that combine visual and textual information. MVNN consistently outperforms the compared visual representations across fusion methods.

  • Multimodal fake news detection: The multimodal experiments fuse textual information with visual representations produced by the evaluated visual modeling methods.
  • Multimodal fake news detection: The comparison uses attRNN, EANN, and MVAE as multimodal fusion methods.
  • Multimodal fake news detection: The multimodal setting evaluates whether MVNN visual representations improve fake-news detection when combined with text information.
  • Multimodal fake news detection: MVNN consistently outperforms other baselines across all evaluated fusion methods.
  • Multimodal fake news detection: Over 5.2% accuracy: MVNN exceeds the compared methods in multimodal fake-news detection.

G. Case Studies

The case studies show that frequency and pixel domains detect complementary evidence in fake-news images, while MVNN combines both to classify examples missed by either single-domain sub-network. The paper also reports MVNN’s effectiveness on the Weibo dataset and identifies broader cross-dataset evaluation as future work.

  • Complementary domain evidence: MVNN combines frequency and pixel-domain evidence to detect fake-news images missed by either single-domain sub-network.The case studies include examples captured by the frequency-domain sub-network and others captured by the pixel-domain sub-network.
  • Frequency-domain cases: The frequency-domain sub-network detects examples whose image semantics show little evidence of fakery but whose frequency histograms appear suspicious.The reported examples were associated with heavier re-compression and possible outdated-image characteristics.
  • Pixel-domain cases: The pixel-domain sub-network detects examples whose frequency histograms show little evidence of fakery but whose visual contents are eye-catching and dubious.The examples include a misinterpreted act of prayer and a mock-up presented as a huge snake.
  • Combined model: MVNN models visual information from frequency and pixel domains to capture fake-news characteristics at physical and semantic levels.The framework dynamically fuses the two types of visual information for fake news detection.
  • Scope and future work: The experiments validate MVNN on Weibo data, while future work proposes testing its generalization on larger and different datasets.The paper states that evaluation was limited to Weibo because of limited distinctive images in the existing Twitter multimedia dataset.
Loading 1908.04472v1…