The next scientific seminar devoted to the topic "Applications of deep learning methods in cyber security: approaches and problems " was held at the Institute of Information Technology of ANAS.

The leading researcher, PhD., associate professor Fargana Abdullayeva, who presented the first report on the topic, highlighted the relevance of the topic and pointed out that one of the innovative and perspective research trends in the application of deep learning methods to ensure cyber security.

She said that despite the fact that deep learning  methods were successfully applied in recognition of images and objects, recognition of speech, and natural language processing, these methods are currently being applied very little to cyber-attack detection. The speaker noted that the existing problems in the field of cyber security, the ability to cope with the growing dynamics of cyber-attacks, failure to detect new threats, difficulties in the process of complex events analysis, and the limited scope of data and attacks, are the main challenges facing the new cyber security solutions. The use of in-depth training methods to offer solutions that will solve these problems is the main attraction for researchers. Deep learning methods have extensive capabilities for successful application of DDoS attacks, behavioral anomalies, malware and protocols, CAPTCHA codes, botnet detection and voice authentication in the field of cyber security.

F.Abdullayeva gave detailed information about the current state of research on cyberbullying methods based on deep network neuron network architectures such as  Deep Belief Network (DBN), Restricted Boltzmann Machine (RBM), Convolutional Neural Network (CNN), Deep Neural Network (DNN), deep autoencoder etc.  basic approaches in these studies, their advantages and problems, and the databases used in experiments to verify the accuracy of the methods.

The head of the department, Ph.D., associate professor  Yadigar Imamverdiyev gave detailed information on discriminative, generative and hybrid neural network architectures based on deep neural networks  and recent trends in the field of in deep learning - Deep Reinforcement Learning and Transfer Learning and  the approach to the application of  Generative Adversarial Networks to detect malicious software: "The discriminative model classifies input data and the generative model is a model that can generate data similar to the training data. In the considered approach, generic model deep encoders learn the characteristics of malware, new data generates and transfers them to teach competitor network generator ". In a discriminative model, the trained discriminator transmits the feature of recognizing malicious software to the detector using transfer learning, the head of the department noted. This approach is also intended to be implemented with variation encoders in future studies.

At the end of the seminar, the views were exchanged and the questions were answered. Y. Imamverdiyev recommended to continue research in this sphere.

© All rights reserved. Citing to www.ict.az is necessary upon using news