This paper presents a transformative cross-modal framework that unifies audio and text modalities through the synergistic use of Convolutional Neural Networks (CNNs) for rich audio feature extraction and BERT (Bidirectional Encoder Representations from Transformers (BERT) for sophisticated, context-aware text embeddings. The proposed framework seamlessly projects both modalities into a shared embedding space, facilitating precise audio-text alignment and retrieval. Unlike traditional models that treat audio and text in isolation, this framework introduces a dynamic CNN architecture tailored to capture intricate audio patterns, while BERT delivers deep semantic understanding of text. Employing cosine similarity loss, the framework ensures optimal cross-modal alignment, achieving high retrieval accuracy and robust generalization across diverse datasets. Tested on the Free Music Archive (FMA) and Speech Commands datasets, the model demonstrated a remarkable ability to cluster and align related audio and text samples with mean similarity scores of 0.82, and achieved cross-modal retrieval accuracy of 0.67, even when applied to domains outside its training scope. This innovative framework offers broad potential for applications such as music information retrieval, speech recognition, and multimedia content analysis, pushing the boundaries of multimodal learning and retrieval systems.
Software vulnerabilities provide openings for cyberattacks and therefore pose a great risk to the security of an IT system. The availability of public exploit scripts strongly influences the scale of the threat as it greatly reduces the effort that adversaries must expend to exploit these vulnerabilities. These exploit scripts also differ relative to maturity, reliability, and skill required to be executed successfully. This factor plays an important role in qualitative and quantitative vulnerability risk assessment, cyber adversary analysis, and cybersecurity training. However, not all public exploit databases offer such a rank, which impairs the accuracy of these evaluations. To address this shortcoming, this study evaluates different machine learning approaches to predict the complexity required to execute an exploit based on the exploit script. For this evaluation, different data preparation approaches and machine learning models, all of which have been proven to perform well on similar tasks, were selected. Each machine learning model was then trained using each data preparation approach to evaluate performance and determine the best approach to reliably predict the required skill to use an exploit. These results can be used to generate complexity predictions for exploits, which can then be included in skill-based assessments and ultimately improve the accuracy of cybersecurity evaluations.
This paper proposes a cutting-edge vulnerability detection method for smart contracts, combining Transformer-based embeddings and Graph Neural Networks (GNNs). Critical opcodes are identified and dynamically weighted using attention mechanisms, enhancing feature representation. The GNN captures both relational patterns and critical opcode characteristics, enabling robust detection of vulnerabilities. Experimental results show significant improvements in F1-scores for both binary and multi-class detection tasks, outperforming traditional models like LightGBM. This approach leverages modern AI advancements to address challenges in accuracy and generalization, providing a scalable and effective solution for smart contract analysis.
Precise segmentation of brain tumors is essential in medical imaging, supporting early diagnosis, personalized treatment planning, and effective monitoring of tumor development and response to therapies. While centralized deep learning models like U-Net are highly effective for this task, they raise substantial privacy concerns when healthcare or-ganizations need to exchange confidential patient information. To address these concerns, the BrainDiffU-Net framework is introduced, a diffusion-based U-Net model designed for decen-tralized learning environments. In this framework, each medical institution independently trains a model on its local dataset and exchanges model updates with neighboring institutions, ensuring data privacy without requiring a central server. The diffusion process allows model parameters to be shared securely across the network, facilitating collaborative learning across multiple institutions while preserving privacy. Simulations conducted on the Low-Grade Glioma (LGG) MRI Segmentation Dataset demonstrate that BrainDiffU-Net achieves segmentation performance comparable to centralized models and significantly outperforms non-cooperative strategies in terms of average loss and Dice coefficient. These results show that BrainDiffU-Net provides an optimal balance between segmentation accuracy and privacy preservation, making it a viable solution for decen-tralized medical applications. The source code is available at https://github.com/Parth-nXp/BrainDiffU-Net.
User satisfaction in Wireless Networks (WLANs) heavily rely on the efficient provision of Quality of Service (QoS). Nevertheless, misbehaving acts by malicious users have the potential to disrupt the QoS provisioning and impair its overall efficiency. One such abusive behavior is called Class Hijacking (CH) and has the ability to deteriorate the QoS-levels received by the well-behaving users in the network. CH is a hard to detect, yet easy to perform attack, targeting the QoS mechanism in differentiated services networks such as WLANs. This article attempts to illustrate the negative impact of CH actions on the network performance of an infrstructure-based WLAN where stealing network resources is an attractive choice for nefarious users. This is achieved through extensive simulations using the ns-3 network simulator. Applying a gradual increase in the number of deceitful users, we observe how these attacks noticeably deteriorate network performance.
