Cloud-based speech recognition services pose privacy risks due to potential exposure of users' personal or sensitive information. To address these privacy concerns, this paper proposes a novel privacy-enhancing method leveraging adversarial attack techniques. Our method applies Gaussian noise-based adversarial examples to input speech signals, obfuscating personal information while preserving speech transcription accuracy. We evaluate the trade-off between privacy preservation and speech transcription accuracy through experiments utilizing the wav2vec 2.0 model for speech recognition and x-vector embeddings for speaker recognition. Experimental results demonstrate that with a 20% noise ratio, the word error rate (WER) remains stable at 0.28, whereas speaker recognition accuracy significantly decreases to 4.5%, effectively enhancing privacy.

Determining the precise worst-case execution time has proven difficult due to the tasks’ increasing complexity (WCET). Inaccurate WCET estimations cause deadline misses, which may result in catastrophic accidents. This can be particularly fatal in the safety-critical systems. In this thesis, a scheduling methodology based on machine learning that can dynamically forecast the WCET of a job has been presented. Since it takes into account both the execution time and job sequence, the proposed method can accurately predict the amount of time needed to finish a task. It employs machine learning to compute the task’s progress rate and allow additional time for execution before the deadline when a task has reached its predetermined bound without being completed, which can prevent a deadline miss caused by inaccurate WCET determination. Additionally, it can reduce the sacrifice of low-criticality tasks by optimizing the HI-mode that guarantees the execution of high-criticality tasks by preempting low-criticality tasks in mixed-criticality systems. This preserves functional safety of the low-criticality task. For the evaluation, C benchmarks and Linux systems were used to investigate the trade-offs between the latency and the accuracy of the proposed technique. The result shows the simple linear regression by ordinary least squares (OLS) with the coefficient of determination (R2) as a percentage was 82%. However, the proposed technique was approximately 99.7% prediction accuracy. This was 3.83% scheduling overhead based on a period of 1 ms in a real automotive task.

Mixed criticality systems have additional safety requirements to prioritize safety-critical functions. However, the design of traditional real-time systems is not satisfactory, and high-criticality tasks may interfere with other low-criticality tasks, resulting in criticality inversion problems such as deadline misses. Although period transformation techniques may be used to solve this criticality inversion problem, an increase in scheduling overhead leads to another problem in which the overall response time of the task increases. In this study, the trade-off relationship between period transformation and scheduling overhead is explained, and an experiment is conducted to analyze the appropriate line of period transformation after reproducing the problem in real-time Linux systems. As a result, it was confirmed that when the period transformation to solve the criticality inversion problem is applied as it is, the overall response time is increased due to an increase in context change by 48.7% and scheduling overhead by 28.7%.

In this paper, we design and implement an Android mobile application that helps hearing impaired people communicate based on STT(Speech-to-Text) and TTS(Text-to-Speech) APIs and accelerometer sensor of a smartphone. This application provides the ability to record what the hearing-Impairment person's interlocutor is saying with a microphone, convert it to text using the STT API, and display it to the hearing-Impairment person. In addition. In addition, when a hearing-impaired person inputs a text using the TTS API, it is converted into voice and told to the interlocutor. When a hearing-impaired person shakes their smartphone, an accelerometer based background service function is provided to run the application. The application implemented in this paper provides a function that allows hearing impaired people to communicate easily with other people when communicating with others without using sign language as a video call.

In mixed-criticality (MC) systems, if a high-criticality (HC) task requires more execution time, the execution time of a low-criticality (LC) task is sacrificed. The MC scheduler conservatively allocates additional execution time for the HC task, exceeding the actual execution time. To give as much as the actual execution time, it is necessary to analyze the execution time pattern of the HC task while minimizing the system load. In this study, we propose MC-LNN, a lightweight neural network (LNN) that minimizes the sacrifice of LC tasks by predicting the actual execution time of HC tasks in a MC scheduler. As a result of analyzing the execution time pattern of the task, MC-LNN showed a 23.03%p higher hit rate than the Ordinary Least Squares method (OLS), which is a linear regression method, for the nonlinear execution time task pattern within 90% of accuracy. Additionally, the proposed technique takes about 0.06% of the average real-time task cycle, 10ms, with a 5.78 μs delay time, showing its usability in real-time systems.

This paper presents the implementation results of an Android application system designed to support communication for individuals with hearing impairments. The system utilizes the Speech-to-Text (STT) API provided by Google Cloud Platform to convert spoken language into text in real time. Additionally, it incorporates Text-to-Speech (TTS) functionality to synthesize speech from text, enabling two-way communication. To enhance usability, the application includes a foreground service that leverages the accelerometer sensor, allowing the app to be launched by shaking the smartphone two to three times. This feature improves accessibility and ease of use for hearing-impaired users in real-world situations.

This paper discusses performance enhancement of a server system by clustering low-end computers. The proposed system connects multiple computers into a single cluster and utilizes Kubernetes to manage load balancing between nodes, thereby maximizing computational throughput and minimizing task execution time. The system dynamically allocates tasks based on memory requirements and the current load status of each node. To evaluate the effectiveness of this approach, Optical Character Recognition (OCR) tasks were executed and the processing speed of a standalone server was compared with that of the clustered system. The results demonstrate that even low-specification computers can be used to build an efficient and high-performing server through clustering.

Disclosed is a predictive maintenance apparatus and method for manufacturing equipment. The apparatus comprises: a communication interface unit configured to continuously receive sensing data generated during the operation of manufacturing equipment over time; and a control unit configured to detect data corresponding to abnormal operating conditions from the received sensing data, predict the remaining useful life of the manufacturing equipment based on the detected abnormal data, and determine an appropriate maintenance schedule accordingly.

Disclosed is an apparatus and method for execution time allocation. The method includes: receiving a plurality of grouped training system calls included in a pre-collected training task; training a prediction model to output the progress level of the training task based on the grouped training system calls; receiving grouped system calls included in a predefined high-priority task; predicting the progress of the high-priority task by applying the grouped system calls to the trained prediction model; calculating additional execution time required for the high-priority task based on the predicted progress; allocating the additional execution time based on the progress of the system calls; and dropping predefined low-priority tasks corresponding to the additionally required execution time for the execution of the high-priority task.