This paper presents a hybrid framework for detecting and classifying image splicing using a combination of deep learning (DL) and traditional machine learning techniques. The proposed method employs Local Optimal Oriented Pattern (LOOP) to extract spatial texture inconsistencies that help distinguish spliced images from authentic images. These LOOP-derived maps are further refined using a pre-trained ResNet-50 model to capture high-level structural features. A Support Vector Machine (SVM) classifier is then used for the final classification between authentic and spliced images.The method was evaluated on publicly available image forensics datasets, including the Columbia Image Splicing Dataset, CASIA v2.0, and a subset of the Dresden Image Database. All experiments used stratified 5-fold cross-validation, and the reported results represent the mean performance across folds. The proposed framework achieved an average classification accuracy of 95.20%. Instead of claiming general superiority, we compare the method with representative feature-based and deep learning approaches using identical evaluation settings. The results indicate consistent detection performance under common post-processing operations. It shows improved tolerance under moderate blur and scaling, while performance decreases under JPEG compression and geometric transformations. This technique also demonstrates low performance variance across validation folds, indicating stable generalization. By combining local texture descriptors with deep residual features, the framework provides a reproducible and statistically supported approach for image splicing detection. The proposed system contributes to digital image forensics by offering a clearly defined evaluation protocol and measurable performance improvements rather than rhetorical performance claims.

Robust multi-object tracking (MOT) is hindered by the instability of appearance features under visual corruptions such as occlusion and motion blur. These perturbations introduce high-variance noise into feature trajectories, weakening temporal representations and causing identity switches. We address this challenge by learning more stable appearance representations under feature corruption. To this end, we propose the Causal Interaction Module (CIM), a causal architecture that follows a filter then reconstruct design for online tracking. A temporal filtering stage summarizes the historical feature trajectory into a stable anchor, and a contextual enhancement stage uses that anchor to refine frame-level features before association. Integrated into standard trackers, CIM improves association robustness while preserving the host tracking formulation. Experiments on multiple MOT benchmarks and corruption stress tests show consistent gains, especially on association-related metrics.

In the Internet of Medical Things (IoMT) field, data sharing is crucial for enhancing the quality and efficiency of diagnosis and treatment. However, due to data privacy and security concerns, data sharing among medical institutions is hindered, presenting challenges in integration and privacy protection. To address these issues, we propose a trustworthy medical data sharing method based on the dual-driven synergy of blockchain and federated learning. Different from previous work, it can resist reasoning, poisoning, and collusion attacks because it covers a more comprehensive discussion on privacy, security, and reputation incentives. This method includes: 1) Privacy protection mechanism: integrating random response and differential privacy technology to resist inference attacks, protect model privacy, and reduce the impact of noise on model performance; 2) Dual-quality threshold aggregation mechanism: Precisely filter out malicious and low-quality nodes through quality thresholds, introduce contribution values including node participation to assign aggregation weights to nodes, improve model performance and resist poisoning attacks; 3) Reputation assessment and incentive mechanism: Calculate the reputation value based on the historical performance of nodes, and design a reputation consensus to encourage honest participation of nodes, punish malicious behavior, and at the same time reduce the entry of malicious nodes into the committee to resist collusion attacks. Finally, we compare our method with state-of-the-art approaches on real-world medical image datasets OrganMNIST_A and BloodMNIST. The results demonstrate that our method achieves superior performance in both Accuracy and F1 Score metrics.

Facial emotion detection has witnessed a surge in demand across numerous applications, including human-computer interaction, healthcare, and security. Accurate expression recognition is crucial for improving human-computer interactions and understanding human behavior. Existing facial emotion detection models face challenges in achieving both high accuracy and real-time processing due to complex architectures. Our goal is to create an efficient yet accurate solution that can work on resource-constrained devices. To address the challenge of accurately recognizing emotions from facial expressions, we propose a novel hybrid approach that combines the strengths of pretrained Lightweight Convolutional Neural Networks (CNNs), and Attention-based Vision Models. The pretrained Lightweight CNN serves as a feature extractor, efficiently capturing facial features, while the at tention model refines the feature representation to focus on crucial regions of the face associated with different expressions. This enables our model to achieve state-of-the-art (SOTA) accuracy with reduced computational re quirements. The proposed model, EmoVisioNet, achieves superior performance across multiple datasets, attaining 99.97 % accuracy on CK+, 96.23 % on RAF-DB, 93.88 % on FER2013, and 96.91 % on FERPlus. The obtained results surpass the current state-of-the-art in this field, demonstrating EmoVisioNet's superior performance in facial expression recognition.

Computer vision is important for improving Intelligent Transportation Systems, especially in a real-time application to detect and identify pedestrians in moving traffic and different weather conditions. This paper introduces an effective detection technique which combines the YOLOv11 model with a Frequency-Focused Convolutional Module (FFCM) as a pedestrian-localising spatial-frequency feature extraction method. The systematic experiments of several variations of the YOLO family-lightweight (YOLO11n) to high-capacity (YOLO11x)-on various urban scenes of CrowdHuman and CityPersons datasets - are conducted. The proposed YOLO + FFCM has been proven to achieve more accuracy, with the highest Average precision (AP) and minimal miss rate, at real-time inference speeds. The experimental results emphasize the performance trade-offs to observe in terms of detection accuracy and computational cost in the various variants of the model. This paper will focus on the applicability of frequency-conscious pedestrian detectors in practical applications and future work is introduced to optimize detection networks under low-light and other limited-resource scenarios.

Face Recognition (FR) is used in a variety of application domains, from entertainment and banking to security and surveillance. Such applications rely on the FR model to be robust and perform well in a variety of settings. To achieve this, state-of-the-art FR models typically use expressive adaptive margin loss functions, which tie the feature norm to concepts related to sample quality, such as recognizability and perceptual image quality. Recently, through the development of Face Image Quality Assessment (FIQA) techniques, biometric utility has become the preferred measure of face-image quality and has been shown to be a better predictor of the usefulness of samples for face recognition compared to more human-centric aspects, such as resolution, blur, and lighting, tied to general image quality. While image quality expressed through feature norms exhibits a certain level of correlation with biometric utility, it does not fully encapsulate all aspects of utility. To address this point, we propose a new adaptive margin loss, FunFace (Face Recognition Through Utility and Norm Estimation), which incorporates biometric utility, estimated by the Certainty Ratio, into the adaptive margin, taking inspiration from AdaFace. We show that FunFace (when used to train a face recognition model) achieves competitive results to other state-of-the-art FR models on benchmarks containing high-quality samples, while surpassing them on low quality benchmarks.

Speech represents a fundamental medium for conveying human emotions and, as a result, speech-based emotion recognition (SER) systems have become pivotal in advancing human-computer interaction (HCI) across a range of applications. While significant progress has been made in speech emotion recognition over recent years, existing solutions still face several key challenges, in that they: $(i)$(i) rely excessively on subjectively annotated (discrete) labels during training, $(ii)$(ii) often overlook the label ambiguity of speech samples that express more than one class of emotions, and $(iii)$(iii) underutilize unlabeled or ambiguous speech, for which typically a label distribution (or so-called soft labels) is available. To address these issues, we propose in this paper a novel SER model that explicitly handles ambiguous speech samples and overcomes the shortcomings outlined above. Central to our approach is a novel real-time soft-label correction strategy designed to refine the annotations assigned to ambiguous speech. The proposed model leverages both, (explicitly) labeled as well as ambiguous samples and applies the dynamic soft-label correction strategy alongside an enhanced inter-class difference loss function to iteratively optimize the label distributions during training. We theoretically demonstrate that our method is capable of approximating the true emotional distribution of speech even in the presence of label noise, suggesting that utilizing ambiguous speech samples without explicit emotion labels still contributes toward more effective emotion recognition. Furthermore, we integrate the representational power of convolutional neural networks (CNNs) with the contextual modeling capabilities of Wav2Vec 2.0 to enable a comprehensive extraction of spatio-temporal speech features. Experimental results on the IEMOCAP multi-label dataset confirm the effectiveness of our approach, achieving state-of-the-art performance with significant improvements in weighted accuracy (WA) and unweighted accuracy (UA) over competing methods.

The rapid development in the field of deep learning has driven the advancement of facial de-identification. It has been widely applied in sensitive areas such as healthcare, social media, and video surveillance. This technology changes facial features to hide personal identities while retaining key details such as expressions and postures, aiming to strike a balance between privacy protection and data utility. However, current methods still struggle to achieve this balance. Generative Adversarial Networks (GANs) dominate the field of facial anonymization due to their powerful image synthesis capabilities. However, transformers and diffusion models, particularly their advantages in representational ability, conditional control, and generation fidelity, remain underexplored. The potential demonstrated by these emerging methods in controllability and anonymization is worthy of further study. This survey systematically reviews latest advancements of deep generative models in face de-identification, analyzes the mechanisms of cross-modal fusion in representative architectures, and proposes a three-level classification framework: generative fusion, deterministic fusion, and adaptive fusion strategies. We focus on current limitations and outline possible directions for future development. We also lays the theoretical foundation and formulate the technical road-map for building a secure and practical facial privacy protection framework, in order to promote the development of privacy security in reality.

Concern about expressing depressive symptoms on social media is growing in the digital age. Traditional detection methods can identify depression but lack clear, human-understandable explanations. We present a novel methodology “InsightDep”, a mental health analysis model specifically designed to detect depression in the social media users. Furthermore, the model provides explainability for the model results and facilitates the translation of the model results and explanations into easily understandable and interpretable formats for human comprehension, leveraging the capabilities of large language models (LLMs). The proposed method extends a variant of the BERT model specifically adapted for analyzing Twitter data, to an explainable and interpretable system. The model makes use of masked attention techniques to perform classification as well as making it explainable. Additionally, we incorporate the power of LLM to convert complex technical explanations into comprehensible comments that improves the interpretability of the system. When evaluated on Twitter and Reddit datasets, InsightDep achieved macro-F1/accuracy scores of 0.599/0.671 and 0.994/0.994, respectively, outperforming other state-of-the-art methods such as BERTweet, TwHIN-BERT, and DepRoBERTa. Our novel methodology focuses on the development of morally aware digital channels that enable prompt action and support for mental health concerns, under the supervision of licensed medical experts like doctors and psychologists. This methodology emphasizes the practical application in therapeutic assessments conducted by these professionals.

Facial emotion recognition models that are used as cognition-oriented signals in biometric pipelines can exhibit demographic disparities even on controlled and realworld datasets. Beyond fairness, such biases can constitute a security vulnerability since subgroup-dependent behavior can be exploited to induce uneven failure modes in downstream decision systems. We propose SynCAF (Synthetic Counterfactual Activations for Fairness), an inference-time framework that mitigates demographic bias without retraining network weights. We do not synthesize new images. Instead, we synthesize counterfactual internal representations by selectively suppressing bias-associated feature maps identified using Integrated Gap Gradient-based Biased Feature Map Suppression (IG2-FMS), to mitigating bias arising from differences in the distribution across various subgroups within the image datasets. By considering the biased feature maps as representative of the sensitive attributes, such as skin color, we suppress them to generate outcomes with reduced ethnic bias. We demonstrate the presence of bias through signed equalized odds (EO) and via visualization using the heatmaps of suppressed feature maps. After implementing the $I G^{2}$-FMS approach, the mean signed EO gap improves from -13.8% to -9.19%, along with a 3.30% improvement in accuracy.

With the continuous advancement of generative models, face morphing attacks have become a significant challenge for existing face verification systems due to their potential use in identity fraud and other malicious activities. Contemporary Morphing Attack Detection (MAD) approaches frequently rely on supervised, discriminative models trained on examples of bona fide and morphed images. These models typically perform well with morphs generated with techniques seen during training, but often lead to sub-optimal performance when subjected to novel unseen morphing techniques. While unsupervised models have been shown to perform better in terms of generalizability, they typically result in higher error rates, as they struggle to effectively capture features of subtle artifacts. To address these shortcomings, we present SelfMAD, a novel self-supervised approach that simulates general morphing attack artifacts, allowing classifiers to learn generic and robust decision boundaries without overfitting to the specific artifacts induced by particular face morphing methods. Through extensive experiments on widely used datasets, we demonstrate that SelfMAD significantly outperforms current state-of-the-art MADs, reducing the detection error by more than 64% in terms of EER when compared to the strongest unsupervised competitor, and by more than 66%, when compared to the best performing discriminative MAD model, tested in crossmorph settings. The source code for SelfMAD is available at https://github.com/LeonTodorov/SelfMAD.

This study presents a novel approach using the Demographic Bias Mitigation Framework (DBMF), which leverages Domain Adaptation (DA) to mitigate demographic biases in affective computing AI systems. Biases in technologies like sentiment analysis (SA) and facial emotion recognition (FER) can result in adverse societal impacts and diminished trust. Unlike typical DA strategies, which adapt data from sources to targets, the DBMF innovatively adapts less biased target data to biased source domains to address gender bias in Natural Language Processing (NLP-SA) and racial bias in Computer Vision (CV-FER) tasks. Statistical methods and fairness metrics confirm the framework’s effectiveness in reducing bias while preserving task performance. Notably, for the CV-FER task, the DBMF achieves state-of-the-art accuracy for facial emotion recognition on a widely used dataset—SFEW2.0—marking a significant advancement in this domain. The framework’s ability to handle diverse tasks, domains, and bias types highlights its potential as a unified solution for bias mitigation. Additionally, the integration of Elastic Weight Consolidation (EWC) ensures the retention of task performance across domains, further reinforcing the framework’s robustness. These findings emphasize the DBMF’s and DA’s critical role in fostering fairness, reliability, and trustworthiness in AI-driven affective computing systems.

Face Recognition (FR) plays a crucial role in many critical (high-stakes) applications, where errors in the recognition process can lead to serious consequences. Face Image Quality Assessment (FIQA) techniques enhance FR systems by providing quality estimates of face samples, enabling the systems to discard samples that are unsuitable for reliable recognition or lead to low-confidence recognition decisions. Most state-of-the-art FIQA techniques rely on extensive supervised training to achieve accurate quality estimation. In contrast, unsupervised techniques eliminate the need for additional training but tend to be slower and typically exhibit lower performance. In this paper, we introduce FROQ (Face Recognition Observer of Quality), a semi-supervised, training-free approach that leverages specific intermediate representations within a given FR model to estimate face-image quality, and combines the efficiency of supervised FIQA models with the training-free approach of unsupervised methods. A simple calibration step based on pseudo-quality labels allows FROQ to uncover specific representations, useful for quality assessment, in any modern FR model. To generate these pseudo-labels, we propose a novel unsupervised FIQA technique based on sample perturbations. Comprehensive experiments with four state-of-the-art FR models and eight benchmark datasets show that FROQ leads to highly competitive results compared to the state-of-the-art, achieving both strong performance and efficient runtime, without requiring explicit training.

Multimodal sentiment analysis aims to understand emotions from text, speech, and video data. However, current methods often overlook the dominant role of text and suffer from feature loss during integration. Given the varying importance of each modality across different contexts, a central and pressing challenge in multimodal sentiment analysis lies in maximizing the use of rich intra-modal features while minimizing information loss during the fusion process. In response to these critical limitations, we propose a novel framework that integrates spatial position encoding and fusion embedding modules to address these issues. In our model, text is treated as the core modality, while speech and video features are selectively incorporated through a unique position-aware fusion process. The spatial position encoding strategy preserves the internal structural information of speech and visual modalities, enabling the model to capture localized intra-modal dependencies that are often overlooked. This design enhances the richness and discriminative power of the fused representation, enabling more accurate and context-aware sentiment prediction. Finally, we conduct comprehensive evaluations on two widely recognized standard datasets in the field—CMU-MOSI and CMU-MOSEI to validate the performance of the proposed model. The experimental results demonstrate that our model exhibits good performance and effectiveness for sentiment analysis tasks.

The widespread dissemination of deceptive content on social media presents a substantial challenge to preserving authenticity and trust. The epidemic growth of false news is due to the greater use of social media to transmit news, rather than conventional mass media such as newspapers, magazines, radio, and television. Humans' incapacity to differentiate among true and false facts exposes fake news as a threat to logical truth, democracy, journalism, and government credibility. Using combination of advanced methodologies, Deep learning (DL) methods, and Natural Language Processing (NLP) approaches, researchers and technology developers attempt to make robust systems proficient in discerning the subtle nuances that betray deceptive intent. Analysing conversational linguistic patterns of misleading data, these techniques’ purpose to progress the resilience of social platforms against the spread of deceptive content, eventually contributing to an additional informed and trustworthy online platform. This paper proposed a Knowledge-Aware NLP-Driven AlBiruni Earth Radius Optimization Algorithm with Deep Learning Tool for Enhanced Deceptive Content Detection (BER-DLEDCD) algorithm on Social Media. The purpose of the BER-DLEDCD system is to identify and classify the existence of deceptive content utilizing NLP with optimal DL model. In the BER-DLEDCD technique, data pre-processing takes place to change the input data into compatible format. Furthermore, the BER-DLEDCD approach applies hybrid DL technique encompassing Convolutional Neural Network with Long Short-Term Memory (CNN-LSTM) methodology for deceptive content detection. Moreover, the BER approach has been deployed to boost hyperparameter choice of the CNN-LSTM technique which leads to enhanced detection performance. The simulation outcome of the BER-DLEDCD system has been examined employing benchmark database. The extensive outcomes stated the BER-DLEDCD system achieved excellent performance with the accuracy of 94 %, 94.83 % precision, 94.30 % F-score with other recent approaches.

In text mining and Natural Language Processing (NLP), extracting emotions from textual data is gaining rapid attraction. The proliferation of online content and the freedom of expression on social media platforms has compelled the advancement of more sophisticated methods for the automatic identification of unwanted content. The One-Dimensional Convolutional Neural Network (1D-CNN) performs best for feature extraction, while the Long Short-Term Memory Network (LSTM) is well-known for extracting sequential features and classification of textual data. This study proposes a novel solution by combining 1D-CNN with LSTM. Utilizing a context-aware model, this study aims to enhance the detection of immoral posts, including hate speech and other forms of immoral content on social media platforms. Various posts are extracted from social media, which are then processed using Glove word embedding models. It depicts the word representations in the text, which yields the form of a real-valued vector determining the words' meaning. The 1D convolution utilizes a filter window over the input time-series data to extract features. Relying on the acquired parameters of the filters, they act as feature extractors. This feature extraction technique allows the model to distinguish detailed patterns and nuances from the textual contents of posts. The sequential processing capabilities of the LSTM network enable an in-depth examination of the text, capturing temporal relationships and context, which is critical for successful content categorization. To validate the proposed model, we utilized a benchmark dataset, called the Twitter dataset. The proposed model achieved the highest accuracy of 95.37

In recent years, generative artificial intelligence has gained momentum and created extremely realistic synthetic multimedia content that can spread misinformation and mislead society. Deepfake detection is a technique consisting of frameworks, algorithms and approaches to predict manipulated contents namely, image, audio and video. To this end, we have analyzed and explored various deepfake detection frameworks by categorizing them as single-modal or multi-modal approaches. To provide better understanding and clarity, single-modal approaches are further categorized as conventional and advanced techniques. Conventional techniques extract complementary handcrafted features and classify them using machine-learning-based algorithms. On the other hand, advanced techniques adopt deep learning and hybrid algorithms to detect deepfakes. Multi-modal techniques utilize a mixture of two or more modalities for feature extraction and fuse them to obtain the final classification scores. These techniques are also categorized either as deep learning or hybrid techniques. The complementary features, multiple modalities, and deep learning models are fused adaptively using score-level or feature-level fusion. The advantages, features, practical applications, and limitations under each category are highlighted to address the challenges and determine future trends to counter deepfakes. In addition, recommendations are also elaborated to evaluate the potential of artificial intelligence in deepfake detection for providing a safer and more reliable digital world.

Speech emotion recognition is of great significance for improving the human-computer interaction experience. However, traditional methods based on hard labels have difficulty dealing with the ambiguity of emotional expression. Existing studies alleviate this problem by redefining labels, but still rely on the subjective emotional expression of annotators and fail to consider the truly ambiguous speech samples without dominant labels fully. To solve the problems of insufficient expression of emotional labels and ignoring ambiguous undominantly labeled speech samples, we propose a label correction strategy that uses a model with exact sample knowledge to modify inappropriate labels for ambiguous speech samples, integrating model training with emotion cognition, and considering the ambiguity without dominant label samples. It is implemented on a spatial-temporal parallel network, which adopts a temporal pyramid pooling (TPP) to process the variable-length features of speech to improve the recognition efficiency of speech emotion. Through experiments, it has been shown that ambiguous speech after label correction has a more promoting effect on the recognition performance of speech emotions.

Patient sentiment analysis establishes an intricate relationship between pain management and sentiment analysis in delivering high-quality medical care. This work presents an efficient pain sentiment recognition system within a smart healthcare framework designed to assess patients' pain levels by analyzing their facial expressions. The proposed system is implemented in four distinct phases. First, facial regions are detected using efficient face-detection techniques. In the second phase, the extracted facial regions undergo feature computation using advancements in deep learning techniques, including end-to-end and pre-trained convolutional neural networks (CNN) to capture complex and discriminative facial features associated with pain emotions. In the third phase, a novel PainCapsule model is introduced, which evaluates pain intensity by analyzing both macro-and microfacial expressions. This phase also employs attention networks, feature tuning, and transfer learning techniques to optimize the system's performance. Finally, in the fourth phase, score fusion techniques are applied to the deep pain recognition models to enhance accuracy and robustness further. The system's effectiveness is rigorously evaluated using two benchmark video datasets: the BioVid Heat Pain Dataset and the Multimodal Intensity Pain (MIntPAIN) database. Extensive experiments and comparative analysis with existing state-of-the-art methods reveal that the proposed system achieves an F1-score of 65.51% for BioVid and 58.31% for MIntPAIN datasets, outperforming other pain recognition systems, demonstrating its potential to advance pain sentiment recognition within smart healthcare frameworks.