10th International Congress on Information and Communication Technology in concurrent with ICT Excellence Awards (ICICT 2025) will be held at London, United Kingdom | February 18 - 21 2025.
Sign up or log in to bookmark your favorites and sync them to your phone or calendar.
Authors - Dominic Gibietz, Daniel Helmer, Eicke Godehardt, Heiko Hinkelmann, Thomas Hollstein Abstract - The ongoing demographic change towards an aging population increases the need for effective solutions to support independent living and ensure the safety of elderly people living alone. Detecting anomalies in the daily routines of these people is a critical task in order to address these challenges and maintain their well-being. This paper proposes an unobtrusive method for anomaly detection using binary sensor data and machine learning. The approach involves a neural network in form of an autoencoder, which evaluates hourly data of each room, including the accumulated residence time, the activity time, and the number of room entries. The system learns individual normal behaviour through online learning and detects deviations from it. Testing and evaluation of the system was carried out using a publicly available dataset and comparing different configurations for the model. A comparison was also made between the use of individual maximum values for each room to normalize the data and uniform values for all rooms, with the former performing significantly better. The results demonstrate that the system can effectively identify the majority of unusual daily routines with a high accuracy, offering potential for improving safety measures for people living alone.
Authors - Chi-Hung Wang, Xiang-Shun Yang, Jun-Yi Liu, Yao-Jun Liu Abstract - Contract review is a common challenge for governments, businesses, and individuals. It becomes challenging when manual reviews are slow, legal expertise is lacking, and clauses are complex. These issues often lead to legal disputes and business conflicts. Traditional rule-based contract review tools often struggle with ambiguous language and unstructured content. Large language models (LLM) can quickly analyze contracts and find risks. But, they are unreliable due to "hallucinations" and a lack of knowledge of rare clauses. This study used retrieval-augmented generation (RAG) technology to overcome these challenges. It integrated verified legal data with large language models. This improved review accuracy to 93.67%. The F1-scores reached 91.95% for compliant clauses and 94.79% for non-compliant ones. The ROC-AUC metric improved to 0.93. The results show that this approach works. It improves the classification and risk identification of contract clauses. It also helps in contract review in the legal and business sectors, promoting the use of legal tech.
Authors - Samrat Ray, Souvik Datta, Smita Mehendale, Mita Mehta Abstract - The use of big data in social justice has become a phenomenon that is transforming the entire society, given that it provides solutions to challenges facing the world through the betterment of the lives of the affected groups of people. This paper focuses on the role of positive change by means of data science with a special emphasis on real-time data analysis in supporting power to the people efforts. It starts with the introduction of Data Science approaches and their connection with social transformation focusing on how it has made it possible for organizations to make sound decisions followed by the practical use of real-time big data to support research claims through the use of real-life case scenarios including poverty alleviation, city planning and development among others. Insights from these shed the light on ethical issues and need to make a conscious effort towards making data science solutions available for every segment of society. Finally, the paper analyses the trends and the future possibilities of data science for social enablement. It highlights the possibility of achieving even greater improvement of social programs through advanced research and development. In conclusion, this paper is a summary of how data science can be used to make society better which should prove useful as a reference for policymakers, researchers, and practitioners who are using data to initiate social change.
Authors - Chi-Hung Wang, Yu-Siang Siang, Yu-Hsuan Lin, Cheng-Hsien Lin Abstract - Aerial imagery is widely employed in intelligent transportation management and urban planning. However, dynamic objects often occlude critical information such as road signs and traffic markings, reducing the accuracy of image analysis and thereby affecting application reliability. Although traditional methods can partially address this issue, their high cost and low efficiency pose challenges in large-scale data processing. To overcome these limitations, this study proposes a background averaging technique based on real-time open-vocabulary object detection integrated with difference-based object detection using depth estimation. This approach enables zero-shot dynamic object removal, enhancing both processing efficiency and scalability. Experimental results demonstrate that our technique outperforms conventional methods across multiple performance metrics. Specifically, the multimodal framework combining depth-based differencing with the YOLO-world model achieves Precision, Recall, and F1-Score of 0.9062, 1.0000, and 0.9508, respectively. Furthermore, the Structural Similarity Index (SSIM) for background reconstruction reaches 0.9603, exceeding that of traditional YOLO models (SSIM = 0.9375). These findings indicate that our method not only effectively removes dynamic objects but also accurately restores background information, providing robust support for applications in intelligent transportation management and urban planning.
Authors - Wa Ode Siti Nur Alam, Riri Fitri Sari Abstract - The rapid development of artificial intelligence has facilitated the creation of Chatbot AI systems capable of addressing diverse healthcare challenges, including public education on critical issues like stunting. Leveraging Generative Pre-Trained Transformer (GPT) models and ensemble learning methods, such systems provide accurate, bilingual responses while ensuring scalability. A key implementation, deploying a Bilingual Chatbot AI through the Telegram application, demonstrates the feasibility of using accessible platforms to disseminate vital healthcare information. However, AI chatbots often face limitations, such as inaccurate or delayed responses, hindering user satisfaction and trust. Challenges in stunting education and nutritional status classification include adapting to linguistic nuances and ensuring real-time interaction. Addressing these gaps, we developed a GPT-Ensemble Learning-based chatbot to deliver information about stunting, including its definition, symptoms, impacts, prevention measures, and classification of toddlers' nutritional status based on gender, age, and height. The chatbots provide relevant responses for stunting education and nutritional status classification in Indonesian and English contexts. Our experiments also highlight Random Forest as the optimal ensemble model, achieving exceptional performance metrics: accuracy (0.99), precision (0.99), recall 0.96, F1-score (0.99), and ROC-AUC (0.99). This high performance ensures reliable nutritional status classification while improving accuracy and speed in bilingual interactions. The results underscore the potential of integrating AI-driven solutions into accessible applications like Telegram, which has significant implications for improving public health awareness and decision-making.
Authors - Andrea Augello, Alessandra De Paola, Domenico Giosue, Giuseppe Lo Re Abstract - Smart systems have become increasingly popular in recent years, widening the attack surface of cyber threats. Machine learning algorithms have been successfully integrated into modern security mechanisms to detect such attacks. Internet of Things (IoT) systems often have limited computational resources and are unable to execute entire machine learning pipelines. However, these systems often produce and manage sensitive data. Thus, it is preferable to avoid exposing their data to external analysis, e.g., on cloud systems. This work introduces SF-AE: a novel architecture that enables the distributed training of an anomaly-based intrusion detection system on devices with limited computational resources without exposing sensitive data. Experimental results on multiple datasets show that SF-AE outperforms state-of-the-art methods in terms of attack detection performance, at lower computation and communication costs for the participating devices.
