Large Language Models (LLMs) are increasingly integrated into sensitive domains such as healthcare and autonomous systems, yet adoption is constrained by security risks that conventional assurance methods do not capture. Traditional software assurance techniques are inadequate for LLM-specific vulnerabilities, including prompt injection, insecure output handling, and training data poisoning. We introduce a quantitative security assurance framework for LLM applications that translates security requirements and vulnerabilities into measurable scores. The framework computes an Assurance Metric (AM) as A M = R M − V M , where VM is weighted using CVSS v4.0, and maps results to five security assurance levels, making security posture comparable, auditable, and actionable. Requirements span input/output validation, training data, development and deployment, access control, third-party services, and security procedures; vulnerability tests align with the OWASP Top 10 for LLMs (prompt injection, insecure output handling, training data poisoning, denial of service, sensitive information disclosure, overreliance, and model theft). Case study results show uncensored models (e.g., Llama2-uncensored) exhibit significantly higher exposure, especially to prompt injection and output-handling attacks–while censored and fine-tuned models attain higher assurance levels. Significance and impact: the framework provides transparent, quantitative scoring to compare systems, prioritize mitigations, and support evidence-based deployment and governance in high-takes environments, with continuous human oversight emphasized.

This work addresses a significant gap in existing image and video databases commonly used in computer vision applications by introducing a unique and comprehensive database named Photo-realistic Multi-Distortion Video Dataset for Benchmarking and Developing Robust Computer Vision Models (PVD4RCV). A key innovation of PVD4RCV lies in its incorporation of some relevant physical factors (e.g. depth information, interaction of light with scene contents) inherent to video signal acquisition in constrained and complex real-world environments, which are used to generate realistic distortions in video sequences (e.g. local motion blur, local defocus blur). PVD4RCV includes a diverse collection of videos featuring common distortions, real-world scenarios, and contextual variations. It includes both original and degraded video versions, along with detailed annotations to support the development of advanced learning models, particularly for tasks such as distortion classification and object detection. This resource aims to advance research and applications in computer vision by providing a robust foundation for model training and evaluation. The database is open source and available at the following link for the community: https://github.com/Aymanbegh/PVD4RCV

The proliferation of social media platforms has significantly contributed to the spread of hate speech, targeting individuals based on race, gender, impaired functioning, religion, or sexual orientation. Online hate speech not only provokes prejudice and violence in cyber-space, but it also has profound impacts in real-world communities, eroding social harmony and increasing the risk of physical harm. This necessitates the urgency for effective hate speech detection systems, especially in low-resource languages such as Norwegian, where limited data availability presents additional challenges. This study utilizes the Barlow Twins methodology, applying a self-supervised learning framework to initially develop robust language representations for Norwegian, a language that is typically underrepresented in NLP research. These learned representations are then utilized in a semi-supervised classification task to detect hate speech. Leveraging a combination of text augmentation techniques at both the word and sentence level, along with self-training strategies, our approach demonstrates the potential to efficiently learn meaningful representations with a minimal amount of annotated data. Experimental results show that the Nor-BERT model is well-suited for detecting hate speech within the limited Norwegian data available, consistently outperforming other models. Additionally, Nor-BERT surpassed all deep learning-based models in terms of F1-score. • Collecting and Annotating Norwegian hate speech data. • Enhanced self-learning with Barlow Twins. • Advancing NLP with Lexical and Semantic Augmentation. • LM-based supervised classification.

The exploration of automated wrist fracture recognition has gained considerable research attention in recent years. In practical medical scenarios, physicians and surgeons may lack the specialized expertise required for accurate X-ray interpretation, highlighting the need for machine vision to enhance diagnostic accuracy. However, conventional recognition techniques face challenges in discerning subtle differences in X-rays when classifying wrist pathologies, as many of these pathologies, such as fractures, can be small and hard to distinguish. This study tackles wrist pathology recognition as a fine-grained visual recognition (FGVR) problem, utilizing a limited, custom-curated dataset that mirrors real-world medical constraints, relying solely on image-level annotations. We introduce a specialized FGVR-based ensemble approach to identify discriminative regions within X rays. We employ an Explainable AI (XAI) technique called Grad-CAM to pinpoint these regions. Our ensemble approach outperformed many conventional SOTA and FGVR techniques, underscoring the effectiveness of our strategy in enhancing accuracy in wrist pathology recognition.

In this paper, we introduced a novel Facial Emotion Recognition (FER) framework that utilizes a diffusion-based approach and an attention mechanism. The model is efficiently trained through self-supervised learning, leveraging labeled and unlabelled data. The proposed framework has been rigorously tested on the FER2013 and AffectNet datasets, achieving promising accuracies of $67.2 \%$ and $68.1 \%$, respectively. The quantitative results not only surpass the performance of existing state-of-the-art FER models but also demonstrate the synergistic effect of combining diffusion-based modeling with self-supervised learning and attention mechanisms within a solid architectural framework. Our approach sets a new benchmark in the field, offering a significant step forward in the accurate and efficient recognition of facial expressions.

ABSTRACT The advent of deep neural networks and improved computational power have brought a revolutionary transformation in the fields of computer vision and image processing. Within the realm of computer vision, there has been a significant interest in the area of synthetic image generation, which is a creative side of AI. Many researchers have introduced innovative methods to identify deep neural network‐based architectures involved in image generation via different modes of input, like text, scene graph layouts and so forth to generate synthetic images. Computer‐generated images have been found to contribute a lot to the training of different machine and deep‐learning models. Nonetheless, we have observed an immediate need for a comprehensive and systematic literature review that encompasses a summary and critical evaluation of current primary studies' approaches toward image generation. To address this, we carried out a systematic literature review on synthetic image generation approaches published from 2018 to February 2023. Moreover, we have conducted a systematic review of various datasets, approaches to image generation, performance metrics for existing methods, and a brief experimental comparison of DCGAN (deep convolutional generative adversarial network) and cGAN (conditional generative adversarial network) in the context of image generation. Additionally, we have identified applications related to image generation models with critical evaluation of the primary studies on the subject matter. Finally, we present some future research directions to further contribute to the field of image generation using deep neural networks.

In this paper, we propose a new comprehensive Video Surveillance Quality Assessment Dataset (VSQuAD) dedicated to Video Surveillance (VS) systems. In contrast to other public datasets, this one contains many more videos with distortions and diversified content from common video surveillance scenarios. These videos have been artificially degraded with various types of distortions (single distortion or multiple distortions simultaneously) at different severity levels. In order to improve the efficiency of the surveillance systems and the versatility of the video quality assessment dataset, night vision CCTV videos are also included. Furthermore, a comprehensive analysis of the content in terms of diversity and challenging problems is also presented in this study. The interest of such database is twofold. First, it will serve for benchmarking different video distortion detection and classification algorithms. Second, it will be useful for the design of learning models for various challenging VS problems such as identification and removal of the most common distortions. The complete dataset is made publicly available as part of a challenge session in this conference through the following link: https://www.l2ti.univ-paris13.fr/VSQuad/.

Web search querying is an inevitable activity of any Internet user. The web search engine (WSE) is the easiest way to search and retrieve data from the Internet. The WSE stores the user’s search queries to retrieve the personalized search result in a form of query log. A user often leaves digital traces and sensitive information in the query log. WSE is known to sell the query log to a third party to generate revenue. However, the release of the query log can compromise the security and privacy of a user. In this work, we propose a Profile Aware ObScure Logging (PaOSLo) Web search privacy-preserving protocol that mitigates the digital traces a user leaves in Web searching. PaOSLo systematically groups users based on profile similarity. The primary objective of this work is to evaluate the impact of the systematic group compared to random grouping. We first computed the similarity between the users’ profiles and then clustered them using the K-mean algorithm to group the users systematically. Unlikability and indistinguishability are the two dimensions in which we have measured the privacy of a user. To compute the impact of systematic grouping on a user’s privacy, we have experimented with and compared the performance of PaOSLo with modern distributed protocols like OSLo and UUP(e). Results show that, at the top degree of the ODP hierarchy, PaOSLo preserved 10% and 3% better profile privacy than the modern distributed protocols mentioned above. In addition, the PaOSLo has less profile exposure for any group size and at each degree of the ODP hierarchy.

Teaching science through computer games, simulations, and artificial intelligence (AI) is an increasingly active research field. To this end, we conducted a systematic literature review on serious games for science education to reveal research trends and patterns. We discussed the role of Virtual Reality (VR), AI, and Augmented Reality (AR) games in teaching science subjects like physics. Specifically, we covered the research spanning between 2011 and 2021, investigated country-wise concentration and most common evaluation methods, and discussed the positive and negative aspects of serious games in science education in particular and attitudes towards the use of serious games in education in general.

Land cover classification has significant contributions in several applications including natural calamities estimation and response, observation of environmental changes, and urban planning to name a few. These types of applications demand the identification of different categories of land cover. Traditional land cover classification is significantly dependent on the availability of huge amount of labeled data. However, labeling satellite data is very time consuming and it often requires expert knowledge. To alleviate the dependency on labeled data, we propose a novel and improved regularization based deep semi-supervised learning (IR-SSL) method for land cover classification. Adaptation of deep semi-supervised learning approach in such a task gains reliability due to its robustness in feature learning. To consolidate the performance of our deep semi-supervised learning method, we combine it with a robust data augmentation technique. We perform experiments on a benchmark dataset. Considering limited labeled samples from the dataset, our method outperforms many state-of-the-art models.

People counting in sports venues is emerging as a new domain in the field of video surveillance. People counting in these venues faces many key challenges, such as severe occlusions, few pixels per head, and significant variations in person's head sizes due to wide sport areas. We propose a deep model based method, which works as a head detector and takes into consideration the scale variations of heads in videos. Our method is based on the notion that head is the most visible part in the sports venues where large number of people are gathered. To cope with the problem of different scales, we generate scale aware head proposals based on scale map. Scale aware proposals are then fed to the Convolutional Neural Network (CNN) and it provides a response matrix containing the presence probabilities of people observed across scene scales. We then use non-maximal suppression to get the accurate head positions. For the performance evaluation, we carry out extensive experiments on two standard datasets and compare the results with state-of-the-art (SoA) methods. The results in terms of Average Precision (AvP), Average Recall (AvR), and Average F1-Score (AvF-Score) show that our method is better than SoA methods.

Human brain undergoes state changes during sleep which produces distinctive signal patterns when recorded by electroencephalography (EEG). Automatic identification of these stages is crucial to diagnosing and treating sleep related disorders. We propose an image processing based technique for automatic identification of sleep stages from EEG signals. We generate two dimensional image representations from the high dynamic range Fourier transform features of the one dimensional EEG signals. Using these representations, we learn a deep and dense convolutional neural network (CNN) model for prediction. The key advantage of the proposed method is its seamless use of the existing well studied and powerful deep CNN models designed for computer vision problems. Experiments on the popular Sleep-EDF database show that the proposed method significantly outperforms the compared methods for automatic sleep stage identification.

People counting in high density crowds is emerging as a new frontier in crowd video surveillance. Crowd counting in high density crowds encounters many challenges, such as severe occlusions, few pixels per head, and large variations in person's head sizes. In this paper, we propose a novel Density Independent and Scale Aware model (DISAM), which works as a head detector and takes into account the scale variations of heads in images. Our model is based on the intuition that head is the only visible part in high density crowds. In order to deal with different scales, unlike off-the-shelf Convolutional Neural Network (CNN) based object detectors which use general object proposals as inputs to CNN, we generate scale aware head proposals based on scale map. Scale aware proposals are then fed to the CNN and it renders a response matrix consisting of probabilities of heads. We then explore non-maximal suppression to get the accurate head positions. We conduct comprehensive experiments on two benchmark datasets and compare the performance with other state-of-theart methods. Our experiments show that the proposed DISAM outperforms the compared methods in both frame-level and pixel-level comparisons.

We propose an End-to-End Transportation Network (EETN) for multi-target tracking. In the EETN, we model the optimal set of trajectories through min-cost flow problem by exploring deep features to generate a graph. The transition cost among the nodes is found through statistical similarity metric. We consider dynamic programming to solve the optimization problem. For experimental evaluation, we compare our proposed EETN method with two state-of-the-art methods using four benchmark datasets. The quantitative analysis shows promising results of our EETN against state-of-the-art methods on precision/recall and F-score.

We propose a novel Hierarchical Feature Model (HFM) for multi-target tracking. The traditional tracking algorithms use handcrafted features that cannot track targets accurately when the target model changes due to articulation, illumination intensity variation or perspective distortions. Our HFM explore deep features to model the appearance of targets. Then, we use an unsupervised dimensionality reduction for sparse representation of the feature vectors to cope with the time-critical nature of the tracking problem. Subsequently, a Bayesian filter is adopted as the tracker and a discrete combinatorial optimization is considered for target association. We compare our proposed HFM against 4 state-of-the-art trackers using 4 benchmark datasets. The experimental results show that our HFM outperforms all the state-of-the-art methods in terms of both Multi Object Tracking Accuracy (MOTA) and Multi Object Tracking Precision (MOTP).

Multi-target tracking is one of the most challenging tasks in computer vision. Several complex techniques have been proposed in the literature to tackle the problem. The main idea of such approaches is to find an optimal set of trajectories within a temporal window. The performance of such approaches are fairly good but their computational complexity is too high making them unpractical. In this paper, we propose a novel tracking-by-detection approach in a Bayesian filtering framework. The appearance of a target is modeled through HoG descriptor and the critical problem of target association is solved through combinatorial optimization. It is a simple yet very efficient approach and experimental results show that it achieves state-of-the-art performance in real time.

In this paper we present a novel method for crowd behavior identification. In our method, the motion flow field is obtained from the video by computing the dense optical flow. Then, a thermal diffusion process (TDP) is exploited to increase the coherence of the motion flow. Approximating the moving particles to individuals, their interaction forces are computed using a modified variant of the social force model (M-SFM) to highlight potential particles of interest. Besides capturing the effect of neighboring individuals on each other, the M-SFM also takes into account the crowd disorder, usually triggered by regions of high interactions. The experimental evaluation is conducted on a set of benchmark video sequences, commonly used for crowd motion analysis, and the obtained results are compared against a state of the art technique.

In this paper we present a novel method for automatic traffic accident detection, based on Smoothed Particles Hydrodynamics (SPH). In our method, a motion flow field is obtained from the video through dense optical flow extraction. Then a thermal diffusion process (TDP) is exploited to turn the motion flow field into a coherent motion field. Approximating the moving particles to individuals, their interaction forces, represented as endothermic reactions, are computed using the enthalpy measure, thus obtaining the potential particles of interest. Furthermore, we exploit SPH that accumulates the contribution of each particle in a weighted form, based on a kernel function. The experimental evaluation is conducted on a set of video sequences collected from Youtube, and the obtained results are compared against a state of the art technique.