Federated learning is a promising framework to train neural networks with widely distributed data. However, performance degrades heavily with heterogeneously distributed data. Recent work has shown this is due to the final layer of the network being most prone to local bias, some finding success freezing the final layer as an orthogonal classifier. We investigate the training dynamics of the classifier by applying SVD to the weights motivated by the observation that freezing weights results in constant singular values. We find that there are differences when training in IID and non-IID settings. Based on this finding, we introduce two regularization terms for local training to continuously emulate IID settings: (1) variance in the dimension-wise probability distribution of the classifier and (2) hyperspherical uniformity of representations of the encoder. These regularizations promote local models to act as if it were in an IID setting regardless of the local data distribution, thus off-setting proneness to bias while being flexible to the data. On extensive experiments in both label-shift and feature-shift settings, we verify that our method achieves highest performance by a large margin especially in highly non-IID cases in addition to being scalable to larger models and datasets.

To fix network congestion resulting from the increase in high volume traffic in data-intensive science and the increase in internet traffic due to COVID-19, there has been a necessity of traffic engineering through traffic prediction. For this, there have been various attempts from a statistical method such as ARIMA to machine learning including LSTM and GRU. This study aimed to collect and learn KREOENT backbone and subscribers’ traffic volume through diverse machine learning techniques (e.g., SVR, LSTM, GRU, etc.) and predict maximum traffic on the following day.

Documents saved in digital libraries have high value as public documents or for commercial use. So, appropriate client authentication and copyright protection systems should be prepared. That is because more and more information is being digitized, as information processing equipment are rapidly progressing, the necessity of security on various kinds of facilities, digital libraries and communications is becoming an important issue. In this study, DLFAS (digital library fingerprint authentication system) is designed and actualized, based on online RTFRM (real-time fingerprint recognition mechanism), centering on fingerprint authentication technology that is decided to be more stable, convenient and credible than password authentication in client identification and authentication using physical characteristics of human. DLFAS that is suggested in this study provides high-quality service effectively in a more stable and convenient environment to authorized clients by solving the weak points and inconvenience of passwords and suggests the model of DLDCPS (digital library digital content protection system) that can prevent unauthorized illegal clients from accessing protected content.

Vast amount of information is generated and shared in this active digital information society. Due to easy access to information in our digital society, there are many cases of illegal counterfeiting and usage of personal information. Producing information with investment and effort is important indeed, but managing and protecting information is becoming a furthermore important issue. This is to promote a new detecting theory and solution for cases of intellectual property violations and plagiarizing digital contents. The document duplication method developed by this research is DICOM (dynamic incremental comparison method) is highly active. It targets on producing various detectors to embody a system to expeditiously and effectively detect duplicated documents and parts.

With the advent of information technology, it is only a matter of time that more and more of digital transactions will be made via the Internet. This increased popularity of the digital transactions has required more efficient protection schemes such as digital rights management system (DRM). DRM protects digital contents, returns profits generated from the contents use to the digital rights providers (DRPs). And it also helps DRPs' gold rush into a new market. This paper introduces a model that contains DRM, and verification methods using applied one-way functional analysis and fingerprinting which is characterized by its immutability and individuality. This model is designed to distinguish clients from unauthorized users, so that only valid users can have access to the digital contents. To sort out unauthorized users, fingerprints recognition mechanism (FRM) is adopted, rather than conventional passwords. The advantages of using FRM surpass the advantages of using passwords. We hope FRMs reliable feature will allow a better digital environment for contents design, distribution and implementation.

This paper compares two clustering methods: SOM, and a graph-based clustering technique , for text-independent speaker verification. The focus of comparison is given to the distribution characteristics of representative frames for each cluster, to the use of processing time of clustering and MLP learning, and to verification performance. Simulation results show that the graph-based technique produces better verification performance than SOM. Other statistics are collected to explain significant difference in MLP learning time with each clustering method. This experiment suggests that there is a best match between a classifier and a clustering method for a given application.

This work addresses the data balancing problem of the existing neural network based speaker verification methods, and proposes new method using modular neural network. In this method, each expert network is trained with the balanced number of genuine speaker data and imposter speaker data. In our experiments, we obtained high performance results for the unknown imposter speakers. High performance and the modular nature of the proposed method enables building a large scalable speaker verification system.