The need for teen safety online is very relevant in the conversation surrounding artificial intelligence. With their limited understanding of AI-based technology, as well as their vulnerability to technological influence and addiction, teens' usage of these tools can become dangerous as they are likely to grow overly trusting of the outputs of fallible AI. Additionally, many AIs do not possess child-safety features. These limitations could lead to users being exposed to misinformation and age-inappropriate content. Coupled with the growing accessibility that teenagers have to AI-powered tools, this creates a need for their safety to be considered when developing these technologies. To address this issue, this paper aims to evaluate the safety of different AI-powered platforms, specifically LLM-based chatbots that are popular among teen audiences, to identify any dangers that a teenager user could encounter when using these tools. The chatbots being tested are ChatGPT, Character.ai, Snapchat's MyAI, and MetaAI, and the conducted tests found all four to be generally safe but identified potential weaknesses of the tools that hinder their overall safety. The goal of this research is to promote the importance of creating an online space that is safe for teenagers and to identify ways AI can become more child safe.

Geographical areas are rapidly changing due to urban development and global change, which are creating difficulties in land cover classification affected by mixed pixels, noise, and spectral variability. To overcome this challenge, this study proposed a novel hybrid classification model that integrates Possibilistic C-Mean (PCM) clustering with learnable spatial filters and uncertainty aware ensemble learning to enhance the classification performance on remote sensing data. PCM is a soft classification method that is known for its robustness and outliers. For generation of land cover maps, PCM is used as a base classifier, and these generated soft classification maps are refined using learnable Gaussian (LGF) and learnable bilateral filters (LBF) whose parameters are optimized using Bayesian optimization to preserve edges and enhance spatial coherence. For improvement in accuracy, we added uncertainty weighted random forest (UW-RF) and uncertainty weighted XGBoost (UW-XGB) classifiers that incorporate pixel-wise uncertainty derived from the previous steps to weight the parameter's influence of each sample during training, where uncertainty is measured using Shannon entropy. The proposed hybrid models PCM+LGF+UW-RF, PCM+LGF+UW-XGB, PCM+LBF+UW-RF, PCM+LBF+UW-XGB are introduced and examined using Landsat 8 images of the Nainital region in India. Experimental results show the improvements in overall accuracy with 97.23% and kappa coefficient of 0.965 while compared to traditional methods. This study highlights the effectiveness of hybridizing the soft classification, adaptive spatial filtering and uncertainty modelling for robust land cover mapping in complex terrains.

As Large Language Models (LLMs) like GPT and Gemini become increasingly integrated into software development, understanding their capabilities and limitations is essential. This project evaluates the effectiveness of these models in code generation by comparing AI-generated code to human-written code in C++ and Python. Key software quality metrics—including cyclomatic complexity, lines of code, and space and time complexity—are used to assess the performance, efficiency, and readability of the generated code. The study also examines how prompt complexity, analyzed at two distinct levels, influences the quality of code produced by the models. By highlighting the strengths and weaknesses of LLMs in handling programming tasks of varying difficulty, this research provides valuable insights for developers, researchers, and industry professionals. The findings aim to inform best practices for integrating AI assistance into development workflows, ensuring a balance between automation and human oversight. Ultimately, this work contributes to more efficient and maintainable coding practices in an AI-augmented development landscape.

Social media platforms such as X (Twitter) generate vast volumes of continuously evolving data, making the extraction of meaningful insights a challenging task. To address this, this paper proposes HDBiTweeC, a novel hybrid clustering framework designed for time-evolving text data. The framework integrates autoencoder-based dimensionality reduction with HDBSCAN, enabling the capture of evolving patterns in data while preserving semantic relationships. For empirical evaluation, tweets corresponding to five trending hashtags were collected using snscrape, pre-processed, and transformed into embeddings. The performance of HDBiTweeC was benchmarked against K-Means, K-Means++, and HDBSCAN using clustering quality metrics including the Silhouette Score, Calinski-Harabasz Index, and Davies-Bouldin Index. Experimental results demonstrate that HDBiTweeC consistently outperforms the baseline methods. In addition, a customized explainability module incorporating z-scoring, polarity-based sentiment analysis, and model-agnostic techniques such as LIME enhances the interpretability of the clustering outcomes. This framework thus enables the discovery of crowd patterns and offers potential applications in identifying emergent events such as disasters, protests, and the spread of misinformation.

Digital forensic investigates evidence from electronic gadgets to aid in explaining criminal activities or security violations. The dynamic nature of technology constantly creates new kinds of digital devices and forms of data, challenging the acquisition and analysis of evidence. A new model named ArithmoGrad Optimization-based Modality Fusion with Attention Network (AG-MFAN) is proposed. ArithmoGrad Optimization (AG) improves deep convolutional neural networks (Deep CNN) by updating the feature extraction processes with arithmetic operations as well as gradients. It also adapts long short-term memory network parameters for temporal and sequential data of audio and documents to enhance the sequence modeling. The Modality Fusion with Attention Network then effectively combines these refined features using an advanced attention mechanism that prioritizes the most relevant information across different data types. This approach addresses the limitations of existing models, resulting in a more effective and precise system for detecting multimedia forgeries. The AG-MFAN model achieves the maximum accuracy, F1-score, precision, and recall of 96.76%, 96.76%, 96.40%, and 97.12% respectively for multi-modal analysis.

The evolution of IoT has brought substantial benefits across several application domains but introduced complex security challenges. Traditional detection systems often struggle to identify attackers more specifically in multiple or integrated attacks. To deal with this problem, this research work proposes an optimized deep fusion network named HydraBoost++ that employs a novel BiConv-FE backbone for feature extraction, MouldSelect in the neck for feature selection, and a boosting head. This overall architecture is inspired by the recent advanced deep learning architectures, and each building block, including backbone, neck, and head, ensures effective handling at each stage of the proposed architecture. The BiConv-FE is mainly designed for the extraction of both local and global context details and to improve classification. However, this integration generates a large amount of information, which is then handled at the neck using the MouldSelect module. The MouldSelect module uses a nature-inspired optimization algorithm and helps in selecting the relevant information for further classification. This stage reduces computational overhead in IoT networks when detection architectures are implemented. The boosting head uses XGBoost as a multi-class classifier and is used to classify the attacker nodes of the network. This proposed HydraBoost++ improves the overall accuracy of the network, which is assessed using publicly available and self-generated datasets. The self-generated dataset consists of labeled simulation traces under different network scenarios and helps in accessing the HydraBoost++ under different network situations. The precision, recall, f-score, and accuracy-based results have shown an incredible contribution of the proposed architecture and have shown better performance than other benchmarked datasets.