Computational researchers have introduced an optimization-guided deep learning framework designed to detect multimodal fake news by coordinating visual and textual evidence while filtering high-dimensional feature noise. Developed to counter deceptive online misinformation that exploits cross-modal camouflage, the system, termed IGWO-MRC, couples an Improved Grey Wolf Optimizer with a multimodal residual collaboration architecture. Across evaluations on a large-scale Chinese multimodal dataset of verified authentic and fabricated news pairs, the architecture achieved a classification accuracy of 92.31 percent, a precision of 94.27 percent, a recall of 90.08 percent, and an F1 score of 92.13 percent. These figures demonstrate an average accuracy improvement of 10.13 percent over competitive baseline models. Modern misinformation frequently disguises deceptive narratives by pairing misleading text with contextual imagery, generating perceptual contradictions that isolated unimodal classifiers cannot reliably detect. By unifying evolutionary swarm optimization with adaptive cross-modal alignment and residual fusion, the framework resolves inter-modal semantic discrepancies while maintaining balanced decision boundaries under noisy, real-world data distributions.

The operational pipeline begins with a structured multimodal feature extraction process engineered to capture complementary structural and contextual patterns from incoming media pairs. Raw textual data undergoes systematic preprocessing, including symbol filtering, stop-word removal, and tokenization, while corresponding image files undergo format standardization and size normalization. To maintain cross-modal integrity, the researchers retained only complete text-image pairs and applied a RandomOverSampler strategy to rebalance category distributions. For the visual modality, the team implemented a two-stream extraction mechanism combining ResNet50 and the Vision Transformer backbone ViT-B/32 from CLIP. Convolutional layers within ResNet50 extract low-level structural details, whereas CLIP generates high-level semantic embeddings aligned with conceptual language representations. Concatenating these vectors produces a unified visual representation. Concurrently, textual sequences are processed through an embedding matrix and fed into a Bidirectional Long Short-Term Memory network. The forward and backward hidden states are concatenated and passed through average pooling, generating sentence-level semantic representations that preserve temporal and contextual dependencies across complex sentences.

To address high dimensionality and redundant representations inherent in deep neural features, the researchers developed an Improved Grey Wolf Optimizer to guide feature selection before multimodal fusion. Standard bio-inspired swarm algorithms, including canonical Grey Wolf Optimization, Bat Algorithms, and Ant Colony Optimization, frequently suffer from premature convergence, limited local search capacity, and inadequate population diversity in high-dimensional spaces. To overcome these constraints, the study incorporated four complementary enhancements into the optimizer. First, the algorithm applies an improved Tent chaotic mapping during population initialization instead of traditional random assignment. By generating chaotic population matrices within bounded continuous intervals and converting continuous parameters into binary selection vectors through explicit threshold rules, the Tent map establishes a broader, more uniform distribution across the search space. This chaotic initialization significantly elevates population diversity from the initial generation, accelerates convergence toward global search corridors, and prevents the optimizer from becoming prematurely trapped in suboptimal local regions during early feature evaluation iterations.

The second and third enhancements within the optimizer govern search trajectory stability and objective balancing. In standard implementations, the alpha wolf parameter remains static, failing to adapt to shifting fitness landscapes. The proposed framework introduces a dynamic adjustment formula where the alpha value scales from 0.5 upward as iterations advance, prioritizing broad feature reduction in early exploration phases while sharpening classification accuracy during later exploitation stages. In parallel, an adaptive fitness function incorporates a dynamic weighting factor to explicitly penalize feature set size against classification accuracy. During initial cycles, smaller weighting values encourage aggressive dimensionality reduction, which gradually transitions to prioritize classification performance as convergence nears. Finally, the optimizer introduces a hybrid search strategy inspired by genetic perturbation. Following position updates guided by the three leading wolves, candidate vectors undergo bitwise mutation with a 10 percent probability through sigmoid thresholding. This stochastic bit-flipping injects controlled exploration diversity, enabling candidate solutions to escape local traps and maintain robust search coverage.

Following feature selection, Principal Component Analysis compresses the retained features into compact representations before they enter the three-stage Multimodal Residual Collaboration backbone. Within this backbone, the Modality-Difference Guided Feature Alignment module resolves structural discrepancies and semantic scale mismatches between heterogeneous visual and textual spaces. Rather than relying on unified semantic projections or shared encoders that allow dominant modalities to overshadow weaker representations, the alignment module employs independent learnable linear mapping matrices for image and text features. These transformations project both streams into a shared d-dimensional embedding space. Once mapped, the module computes a cross-modal coupling factor via normalized cosine similarity between the projected vectors to quantify semantic consistency. Utilizing this consistency indicator, the framework derives a modality-difference regulation factor through learnable weight transformations of the subtracted feature vectors. This gating mechanism adaptively weights image and text components to generate a stabilized aligned collaborative representation without depending on computationally demanding explicit attention maps or complex multi-head interaction layers.

To prevent dominant modalities from being diluted during synthesis, the aligned features are processed through a Squeeze-and-Excitation channel attention block before entering the Residual Gated Fusion module. Modality contributions often vary significantly across distinct news samples, creating representational asymmetry where conventional static concatenation or simple attention weighting leads to feature distortion. The residual gated structure models cross-modal information coordination across four interconnected steps: semantic mapping using hyperbolic tangent activations, implicit consistency modeling, gated weighted path synthesis, and dominant modality residual compensation. The network computes contribution gating ratios through sigmoid activation across concatenated representations, selectively blending visual and textual channels. Simultaneously, the framework establishes a dedicated visual residual pathway that routes normalized image features through batch normalization, non-linear rectified linear unit filtering, and a learnable transformation matrix. Scaled by a designated residual control factor, this compensation signal augments the central gated fusion output, preserving the structural stability of dominant visual evidence and preventing informational collapse during challenging cross-modal interactions.

To sharpen decision boundaries and ensure clear class separation in the latent embedding space, the fused representations are supervised by a Contrastive Modality Clustering Loss module paired with Focal Loss. In real-world environments characterized by pseudo-paired samples or semantic drift, fused features frequently exhibit distributional overlap between authentic and fraudulent classes. The contrastive module constructs image-specific and text-specific views from the fused output, establishing a bidirectional supervised constraint across both image-to-text and text-to-image directions. Governed by a temperature hyperparameter and label-driven sample pairing within mini-batches, the objective encourages representations belonging to identical veracity classes to cluster together while repelling discordant pairs. Furthermore, an inter-class margin control mechanism enforces a defined semantic distance between positive and negative pairs to enhance structural robustness against ambiguous samples. This contrastive regularizer is combined with Focal Loss, which dynamically downweights easily classified instances and focuses training on hard negative examples, yielding a joint optimization objective that mitigates severe class imbalance and noisy labeling.

Empirical validation of the framework utilized a newly constructed multimodal dataset gathered from Sina News, Science China, and the Joint Internet Rumor Refutation Platform. The corpus contains 7,538 news texts and 6,262 accompanying images. Sina News provided 2,449 verified authentic texts and 1,260 images covering current affairs, finance, and technology. Science China contributed 4,186 fabricated texts and 4,100 images targeting scientific communication and pseudoscience debunking, while the rumor-refutation platform supplied 903 fake texts and 902 images. The data was partitioned into 56 percent training, 14 percent validation, and 30 percent testing subsets. Preliminary unimodal benchmarks demonstrated the performance limits of single-stream architectures. Among text models, BiLSTM attained 87.43 percent accuracy and 88.11 percent precision, but suffered from 63.67 percent recall, while TextCNN and TextRNN achieved lower precision. For visual models, MobileNetV2 reached 88.96 percent accuracy and 79.09 percent precision, whereas ResNet achieved an F1 score of 69.06 percent, confirming that unimodal systems exhibit substantial coverage bottlenecks.

In comparative evaluations against established multimodal architectures, IGWO-MRC demonstrated consistent superiority across all primary evaluation metrics. The framework achieved 92.31 percent accuracy, 94.27 percent precision, 90.08 percent recall, and a 92.13 percent F1 score, outperforming baseline models including MM-FakeBuster, which attained an 88.84 percent F1 score, and EANN, which achieved an 88.52 percent F1 score. Other baselines, such as SAFE, MVAE, ERIC-FND, and MDFEND, demonstrated lower precision or reduced accuracy, reflecting difficulties in managing false positive rates and cross-modal semantic discrepancies. When evaluating optimizer performance on the identical multimodal classification backbone, IGWO-MRC surpassed Particle Swarm Optimization, canonical Grey Wolf Optimization, Bat Algorithms, and Harris Hawks Optimization. Furthermore, mathematical convergence assessments across four standard CEC benchmark test functions—Sphere, Rastrigin, Rosenbrock, and Griewank—validated that the improved optimizer maintains superior descent trajectories, faster convergence velocities, and enhanced resistance to local optima entrapment compared to nine competing metaheuristic optimization algorithms under identical iteration limits.

Comprehensive ablation experiments confirmed the structural contribution of each modular element within the architecture. Removing the BiLSTM sequence encoder caused the most pronounced degradation, reducing accuracy to 77.11 percent and F1 score to 78.71 percent. Eliminating the Residual Gated Fusion module reduced the F1 score to 82.25 percent, while removing the Squeeze-and-Excitation channel attention block lowered accuracy to 85.95 percent. Disabling inference-stage adaptive thresholding dropped precision to 78.17 percent, and replacing Focal Loss with standard cross-entropy decreased the F1 score to 83.39 percent. Position sensitivity testing indicated that deploying feature selection before fusion was critical, as post-fusion selection caused severe performance drops. On the public Weibo benchmark, the architecture achieved 90.8 percent accuracy and an 89.5 percent F1 score, outperforming large language model baselines including Zero-Shot, Few-Shot, Z-CoT, and LeRuD. The authors noted that the model operates on static content without temporal evolution tracking and highlighted scalability to larger multimodal corpora as an ongoing research objective.

Subject of Research: Multimodal fake news detection using swarm intelligence optimization and residual feature fusion

Article Title: IGWO-MRC: optimization-guided multimodal fake news detection via residual collaboration

Article References: Mu, G., Dai, J., Ma, X., Liu, Z., & Li, J. (2026). IGWO-MRC: optimization-guided multimodal fake news detection via residual collaboration. EPJ Data Science, 15(1), Article 84. https://doi.org/10.1140/epjds/s13688-026-00681-y

Image Credits: AI Generated

DOI: 10.1140/epjds/s13688-026-00681-y

Keywords: Multimodal fake news detection, Grey Wolf Optimizer, Residual gated fusion, Cross-modal alignment, Contrastive learning, Deep learning, BiLSTM, Feature selection, IGWO-MRC, optimization-guided, multimodal, fake