We are exploring how large language models (LLMs) can generate course outlines and how iterative prompting can refine them. A key challenge is that models often repeat ideas instead of adding new ones, which creates extra manual review. To address this, we developed a novelty metric that combines lexical similarity with semantic analysis to classify content as either new or repeated. Our findings show this metric also helps compare model performance in iterative prompting.
With social media now a main source of news, the spread of misinformation has become a serious concern. We focus on detecting fake news using fine-tuned large language models (LLMs). By training both open-source and proprietary models such as RoBERTa and ChatGPT on a large news dataset, we significantly improved their ability to classify articles as real or fake. Evaluated through accuracy, precision, recall, and F1-score, our results show fine-tuning greatly boosts performance. This work highlights the potential of LLMs to provide reliable AI-driven solutions that help combat misinformation and strengthen trust in online news.
We conducted a comprehensive three-fold study aimed to address the privacy and security of information stored in ChatGPT's memory. Our work involved 135 users who participated in a survey asking them to share the information stored by ChatGPT's memory feature in their user accounts. We utilized respondents' answers to assess their awareness, attitudes, and behaviors regarding ChatGPT's memory feature using the Knowledge-Attitude-Behavior (KAB) model. In addition, we analyzed the information stored in their accounts to evaluate the performance of ChatGPT's memory algorithm in extracting accurate and relevant information without potentially compromising users' privacy and security. Finally, we determined the distribution of stored data across key categories (e.g., personal interests, education, preferences, contact information, etc.) to gain insights into the nature of the information retained by the system.
We implemented and analyzed the 5G NR V2X open source module in network simulation-3 (ns-3) and then demonstrated the NR V2X performance in terms of different key performance indicators by varying a number of situations, including transmit power, communication range, packet size and sub-carrier spacing in the broadcast channel.
Video analytics applications often require fast response and high accuracy. Performing the video analysis at the network edge can provide significantly lower latency and also reduce bandwidth consumption. Due to these advantages, video analytics in edge computing environments have been gaining popularity. In our pursuit to study video quality from the perspective of video analytics tools, we will design and implement a quality-aware video analytics platform based on edge computing.
Publication: [1].
With the prevalence of wireless imaging applications, there is an emerging need to support the exchange of videos among different wireless communication entities. Software-defined radio (SDR) is a promising technology for handling the needs of video communication in dynamic wireless conditions.
We designed a cross-layer video streaming mechanism over SDR to maximize the perceptual video quality. First, we built an SDR platform comprising of GNU radio, USRP B210 and VERT2450 antennas, and GStreamer. Then, we analyzed the experimental results from the platform to determine how the parameters in the physical and the application layer influence the QoE. Based on these results, we formulated a streaming algorithm that simultaneously adjusts the video resolution, bitrate, frame rate, Tx gain, and frequency to maximize the QoE.
Our SDR video streaming platform won the Best Team Project in the Constrained category at the 2020 National AFRL Beyond 5G Software Defined Radio (SDR) Challenge.
Content-centric networking (CCN) is a future Internet architecture that focuses on the content or data rather than its location. It offers more efficient and scalable communication than traditional host-to-host networking solutions. The delivery of multimedia information in CCNs requires sophisticated design due to the large amount of traffic, complicated syntax and bitstream organization, and quality requirements from users.
We developed a novel QoE-aware multisource video streaming scheme for CCN. First, we studied the content distributions of video files among CCN nodes for different caching methods. Then, we designed an adaptive video streaming with distributed caching (ASDC) algorithm to guarantee satisfactory QoE. The ASDC algorithm considers the delivery of scalable video streams and automatically adapts the layers in a video stream based on a QoE model that characterizes the effect of stalling.
UAVs are used in many surveillance and remote sensing applications because of their ability to fly with sensing devices over large areas. The sensed data such as weather information or surveillance footage is usually very time sensitive, and a successful mission requires the data to be captured and delivered to processing centers in time. So, the data collection and offloading strategies of UAVs should be carefully determined to ensure the timely delivery of information while saving the travel costs.
We designed a novel path planning strategy for UAVs that finds paths with minimal cost to meet application deadline requirements under time window and communication constraints. First, we formulated a Mixed Integer Linear Programming (MILP) based path planning problem considering realistic data ferrying scenarios that include the complete mission of UAVs from data collection to data offloading, in which operation time windows and variable channel bandwidths are considered for each waypoint. Then, we designed a genetic algorithm (GA) based algorithm to solve the optimization problem.
Publication: [1].
Smart energy metering systems are in high demand as they offer quick revenue collection, remote monitoring, and control of power distribution system. We proposed, designed, and implemented a low-cost universal smart energy meter (USEM) with demand-side load management. The meter can be used in the postpaid and prepaid modes with flexible tariff plans such as time of use and block rate tariff.
The smart meter comprises of a potential transformer, current transformer, and microcontroller unit with an embedded communication module. The connectivity among the utility authority, the smart meter, and consumer is established by exchanging identification numbers via the cellular network. The load management option of the meter controls electrical loads and provides emergency power during the power shortage. The USEM can be configured and reconfigured remotely by SMS. Moreover, energy consumption status, meter tampering, and fault at the distribution end can be monitored with the proposed metering system.
Publication: [1].
Despite being the 5th most-spoken native language and the 7th most spoken language by total number of speakers in the world, there are very few efficient Bangla OCR systems. The number of systems that can deal with graphical elements are even fewer.
We built a Bengali OCR framework based on Texture-based Image Segmentation that works on printed Bengali text. The framework uses five Gray Level Co-occurrence Matrix (GLCM) features and structural analysis for image segmentation. Then, our proposed Normalized Cross Correlation (NCC)-based algorithm is employed to retrieve the Bengali text from document image.
More details: [1].