The Artificial Intelligence Smart Classroom Initiative (ASCI) presents a re-imagined set of online course tools, designed primarily to support growing computer science classes. The system has four primary tools: an office hours queue, an automatic student grouping algorithm, a course-specific local large-language model (LLM), and administration tools for detecting students and TAs that need support. These tools interoperate to improve the quality of one another (e.g., LLM conversations support students directly in the office hours queue) and are enhanced by synchronizing data from multiple external sources such as Piazza, Gradescope, and Canvas. The system has been deployed in multiple courses over the past three semesters: initially as a FIFO queue, then supporting manual grouping and smart grouping of office hour attendees, and recently including LLM support. Preliminary results indicate that students who were grouped using the tool were more likely to return to the queue more than twice as often (on average) than those who were not. However, while grouping in office hours has the potential to decrease student wait times, teaching assistants and students tend to favor one-on-one meetings over group meetings. This might be improved in the future with updates to the software, TA training, and incorporation of other supporting tools (e.g., LLM technology). The other, newer, tools will be more thoroughly evaluated in future semesters.

As undergraduate enrollment in computer science rises, instructors continue to investigate methods to improve the student experience at scale. One aspect commonly used in courses at scale is queue-driven office hours, in which students join an online queue and meet with teaching assistants on a first-come, first-serve basis (FIFO).

Online assessments allow instructors to facilitate exams and quizzes in both virtual and large classes. Having a clear online timer during these assessments is vital to help students manage their time. However, these same timers can be a cause of anxiety, affecting student performance. Our goals were to determine (i) which types of visualizations are currently in use, (ii) which styles of online timer were preferred by students, and (iii) if providing students a choice of timer impacted their performance.

Visualizing time limits during online assessments is a cause of anxiety, affecting student performance. An initial survey of 34 students across two Computer Science courses found that time-tracking devices produced anxiety for 67.7% of students. While students differed on timer color preference, a majority preferred a count-down display showing time remaining with the ability to hide the timer. In a small pilot study across five exams, we employed multiple time-tracking displays. Preliminary data suggests that students presented with a count-down grayscale timer performed better on average than those presented with a green-yellow-red (GYR) version. Other displays, such as text-only digital count-down timer or elapsed time progress bars, did not elicit as large a difference in performance. These findings indicate the need for further study.

We present a web application system where users can learn about and practice SQL injection attacks. Our system is designed for students in a university level database or computer security class, and is aimed towards students familiar with SQL but with little experience in web security. Our platform currently contains 12 levels, each of which demonstrates a SQL vulnerability that the user must exploit. For each level, we explain the goal of the challenge, and also provide detailed solutions. Our system provides advantages over other methods of teaching SQL injection because it is hands-on, the challenges provide a greater scope of vulnerability coverage, and is easily extensible, allowing instructors to add their own SQL injection problems for their students.

Faced with the continuously increasing scale of data and expectation on response time, complex deep learning technologies, though highly accurate, present two non-rival challenges: a large amount of training data makes a model impossible to be built in short time and intolerable time-cost prohibits acceptable real-time responses. In this research we focus on improving the accuracy and efficiency of the handwritten digit recognition problem. We chose this problem because it is regarded as the prototype of a lot of complex recognition and classification problems. The success of classification of the handwritten digit dataset can be extended further to other advanced areas. The Convolutional Neural Network (CNN) is implemented to do the recognition. We further improved the accuracy by adding elastic distortion to the input data, which helps the model better select the features. In addition we implement distributed computing to reduce the time cost. The training process is divided and a final model is formulated by the combination of each trained model. The results shows two facts: the elastic distortion helped the CNN model to improve the accuracy by 7-10%; and the distributed computing method reduced the training time consumption by about 50%.

Internet traffic measurement and analysis have long been used to characterize network usage and user behaviors, but face the problem of scalability under the explosive growth of Internet traffic and high-speed access. In this paper, we present Hobbits, a Hadoop and Hive based traffic analysis system that performs Internet Protocol (IP) and Transport Control Protocol (TCP) analysis of large-sized Internet traffic in a scalable manner. Our experimental evaluation on real datasets confirms that Hobbits outperforms previous solutions in terms of both job completion time and storage efficiency.