DL methods such as deep autoencoders, limited Boltzmann machines, recurrent neural networks, generative adversarial networks, and others are described in a tutorial-style manner. Each of the DL techniques is then discussed in terms of how it is utilized in security. For example, intrusion or virus detection, and biometricbased user authentication, machine learning (ML) has been extensively used in cybersecurity during the last several years. ML algorithms, on the other hand, are susceptible to assaults during both the training and testing stages, resulting in significant performance drops and security breaches. Studies on the vulnerability of machine learning methods against security threats and associated protective measures are rare. Some wellknown datasets are also included since they are an essential component of ML methods. When to utilize a particular algorithm is also discussed. MODBUS data from a gas pipeline was used to evaluate four machine learning methods. ML algorithms were used to classify a variety of assaults, and the performance of each method was then evaluated.

It is no secret that Machine Learning as a Service (MLaaS) cloud platforms are becoming more popular as machine learning (ML) and deep learning (DL) methods improve. Outsourcing the training of Deep Learning (DL) models is increasingly being done via third-party cloud services, which need expensive computing resources (such as graphics processing units (GPUs), for example). Cloud-hosted ML/DL services are so widely used that attackers have a broad variety of attack surfaces from which to exploit the system. An assessment of cloud-hosted ML/DL models in terms of assaults and defenses is conducted in this paper. Security risks have been identified against a number of learning techniques in the past, including naive Bayes, logistic regression, decision trees and support vector machines (SVM). Due to the nature of the danger, we examine it from two different perspectives, namely the training and testing/inferring phases. Defensive machine learning methods are then divided into four categories: security assessment mechanisms,https://www.ijaer.com/admin/upload/06%20N%20Lingareddy%2001165.pdf countermeasures in the training phase, countermeasures during testing or interpretation, data security, and privacy. Last but not least, we highlight five noteworthy themes in the research on machine learning security risks and protective methods, which need further study in the future.

The evaluation of an intelligent Humanoid Robot system is not very common. These intelligent robots answer the query without the interference of human. This research proposes a Humanoid Robot with oneself learning ability to acknowledge and react to individuals dependent on Deep Learning and enormous information from the web. These types of robot generally used in hotels and resorts, educational institution and the public sector. The Humanoid Robot ought to think about the style of inquiries and close the appropriate response through discussion among robot and client. In our condition, the robot will recognize the clients face and acknowledge orders from the client to do an activity. The inquiry from the client will be handled utilizing profound learning and will contrast the outcome and information on the framework. Our research used GRU/LSTM, CNN and BiDAF with massive SQUAD data set in Deep Learning methods for training purposes. Our research shows that implementing BIDAF with GRU/LSTM encoder gives higher accuracy matching and F1 Score than the BIDAF model with CNN.