Identifying emotions and arousal levels from voice data has gained tremendous traction due to its broad applications in fields such as human-computer interaction, mental health assessment, and education. In this work, we combine Wavelet transform and Mel-frequency cepstral coefficients (MFCC) for preprocessing audio data to recognize emotions and arousal levels from speech. This approach is applied to a convolutional neural network (CNN) model, which classifies a range of emotional states and arousal intensities.

With video data being prevalently used in numerous applications, there is a substantial privacy risk due to the potential exposure of personal information embedded in these video datasets, especially facial features. In this work, we propose a method for concealing personal facial features within video content. It aims to safeguard individuals’ personal privacy by preventing inadvertent exposure and ensuring their anonymity. This method performs video content analysis, including scene detection, face detection, and tracking, and safeguards the personal privacy of individuals by applying face masking techniques, except for target individuals of interest. We have demonstrated the superior performance of our proposed method, as evidenced by the reduction in facial exposure scores. This indicates the method’s ability to consistently uphold privacy across video sequences while minimizing the unintended exposure of personal information.

In this work, we introduce a novel approach for addressing class imbalance in long-tail datasets, focusing on the strategy of dividing and merging. Our methodology involves dividing the dataset into separate clusters, each representing a different segment of the class distribution. By training dedicated models for each cluster and subsequently merging their parameters, we achieve a more balanced learning process across diverse class sizes.

Artificial intelligence (AI) based intrusion detection system (IDS) leverages machine learning and deep learning algorithms to analyze and identify patterns that may indicate malicious behavior. Achieving real-time operation in AI-based IDS involves a tradeoff between inspection performance and the complexity of feature extraction, influenced by the duration of traffic flow observation. To reduce the delay in feature extraction, we developed an AI-based IDS with parallel traffic feature processes utilizing off-the-shelf computing devices. The proposed AI-based IDS can process traffic flows and conduct inspections at a rate of 2,800 flows per second.

Deep learning-based cybersecurity applications should be able to continually accumulate threat knowledge for new types of threats over time while maintaining the knowledge of threats already exposed to the application. This paper proposes episodic memory management for continual learning with network intrusion datasets. For new attacks, the number of samples may not be sufficiently large for training, and thus the memory management algorithm should retain as many samples as possible instead of random sampling in the episodic memory for continual learning. The experiment results indicated that the proposed algorithm outperforms offline learning in terms of average per-class accuracy in a continual scenario with a network intrusion dataset.