Custom Binary Cross Entropy Based Anomaly Detection in Bank Transactions using Deep Convolutional Neural Network

Kulkarni, Chinmay; Mohith,; Sharath, Shylaja S

doi:10.1145/3487664.3487708

Cited by 1 publication

(1 citation statement)

References 2 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…where f (w, x n,k , r n,k ) captures the error of the model parameter w for the input-output vector pair {x n,k , r n,k }. Since we focus on binary health risk assessment, the function f (w, x n,k , r n,k ) used in our model is the binary cross-entropy, which is highly recommended for binary classification problems [15]. The training process is capable to find the global parameter w which minimizes the loss function on the whole data set, which is given by J [11].…”

Section: Developed Fl-based Toolmentioning

confidence: 99%

Health Risk Assessment with Federated Learning

Diamantoulaki

Diamantoulakis

Bouzinis

et al. 2022

2022 International Balkan Conference on Communications and Networking (BalkanCom)

View full text Add to dashboard Cite

The latest growth of storage capabilities has led to an accumulating volume of medical data stored locally by various healthcare entities. Given the recent progress observed in the domain of artificial intelligence, these data could be efficiently exploited, leading to improved and less expensive healthcare conditions. However, the common practice is medical data to be solely locally used, ending up poorly exploited, due to strict sharing restrictions stemming mainly from privacy limitation of their sensitive nature. Considering the centralized character of conventional machine learning approaches, it is apparent that they cannot reassure the privacy required. On the other hand, federated learning (FL) can be regarded as an upcoming and promising answer to efficient exploitation of medical data, considering its decentralized approach. In more details, FL can enable collaboration among various participating entities on the development and training of a common, central and fully shared model without need of sharing owned sensitive data. Thus, apparently FL approach not only can mitigate the privacypreservation issues but can lead to the development of reliable and robust healthcare tools. Indicatively, FL can facilitate the development of a model capable of assessing the health risk, which can be a vital tool for medical sciences. To this end, in this work we present a tool capable of assessing the occurrence of different diseases or complications. This tool is based on FL technique utilizing deep neural network model. The FL model developed herein is indicatively applied to four different medical applications proving its generality in the healthcare domain. The FL approach discussed herein is compared with a corresponding centralized learning. According to the demonstrated results, FL can consist a useful health risk assessment tool exhibiting acceptable performance while preserving privacy in sensitive medical data.

show abstract

Section: Developed Fl-based Toolmentioning

confidence: 99%