Search

Audio deepfakes have been increasingly emerging as a potential source of deceit, with the development of avant-garde methods of synthetic speech generation. Hence, differentiating fake audio from the real one is becoming even more difficult owing to the increasing accuracy of text-to-speech models, posing a serious threat to speaker verification systems. Within the domain of audio deepfake detection, a majority of experiments have been based on the ASVSpoof or the AVSpoof dataset using various machine learning and deep learning approaches. In this work, experiments were performed on a more recent dataset, the Fake or Real (FoR) dataset which contains data generated using some of the best text to speech models. Two approaches have been adopted to the solve problem: feature-based approach and image-based approach. The feature-based approach involves converting audio data into a dataset consisting of various spectral features of the audio samples, which are fed to the machine learning algorithms for the classification of audio as fake or real. While in the image-based approach audio samples are converted into melspectrograms which are input into deep learning algorithms, namely Temporal Convolutional Network (TCN) and Spatial Transformer Network (STN). TCN has been implemented because it is a sequential model and has been shown to give good results on sequential data. A comparison between the performances of both the approaches has been made, and it is observed that deep learning algorithms, particularly TCN, outperforms the machine learning algorithms by a significant margin, with a 92 percent test accuracy. This solution presents a model for audio deepfake classification which has an accuracy comparable to the traditional CNN models like VGG16, XceptionNet, etc.

The quantity and quality of the training data greatly influence the outcome of deep learning-based face-related tasks. However, gathering and labeling sufficient samples of adequate quality with a balanced distribution remains an expensive and laborious task. Hence, several data augmentation approaches have been extensively employed to enhance the quality and increase the quantity of the training dataset. This paper proposes a deep learning Convolutional Neural Network (CNN) based approach for Deepfake detection on the CelebDF dataset, in which several data augmentation techniques such as traditional augmentations, GridMask, face morphing, and style transfer are systematically examined to improve the capabilities of state-of-the-art Deepfake detection approaches. The problems and positive effects of each augmentation method are analyzed and compared in a concise yet informative manner. The trained model is then evaluated against an unseen test datum based on metrics like precision, recall, and F1-score, thereby deciding the best augmentation strategy for the chosen dataset.

In this paper, a honeypot system has been presented, which conducts a severity analysis of the adversaries who attack it. The Honeypot systems are deployed by various organizations to protect their real systems from external threats. They consist of fake file-systems that remain aloof from the attackers. Honeypots gather logs of the attacks to protect the genuine systems from attackers. However, attackers also deploy honeypot detection tools. To defer detection from the attackers, a Q-learning based on an SSH-based honeypot named Cowrie has been implemented to make it adaptive and obtain as much information as possible about the intruder. Severity analysis has been implemented to classify attacks based on their severity. This can be used by real systems to enhance their firewalls, Intrusion Detection Systems, and other security mechanisms against these threats.

Accurate magnetic-signature prediction is the most challenging task in the path of active magnetic-signature minimization. This article discusses the effect of the change in hull thickness, relative permeability, and radius of the object, considering the detailed understanding of the magnetic signatures. The traditional magnetic experiment-based method using a 12-coil structure is implemented for a double hull vessel. However, the time required and effort involved in predicting the magnetic signature is much higher. Hence, a multiple linear regression-based statistical machine-learning (ML) technique for the magnetic signature prediction is proposed. The nonlinear magnetic behavior of the naval vessel is assumed to be linear, because the hysteresis curve in the range of the earth’s ambient field is linear. Hence, a linear relationship is developed between the magnetic signatures and each component of the ambient earth’s magnetic field. The ML-based model also considers the crosstalk among the longitudinal, athwartship, and vertical signatures. The predicted magnetic signatures using the ML technique and the experiment-based techniques are compared. The ML method is found to be accurate, fast, and precise compared with the laboratory-based method, which has strict and continuous requirements for hardware magnetic facility. This method is very simple to implement and directly contributes to the development of the real-time degaussing current calculation algorithm.

The usage of the internet as a fast medium for spreading fake news reinforces the requirement of computational utensils in order to fight for it. Fake videos also called deep fakes that create great intimidation in society in an assortment of social and political behaviour. It can also be utilized for malevolent intentions. Owing to the availability of deep fake generation algorithms at cheap computation power in cloud platforms, realistic fake videos or images are created. However, it is more critical to detect fake content because of the increased complexity of leveraging various approaches to smudge the tampering. Therefore, this work proposes a novel framework to detect fake videos through the utilization of transfer learning in autoencoders and a hybrid model of convolutional neural networks (CNN) and Recurrent neural networks (RNN). Unseen test input data are investigated to check the generalizability of the model. Also, the effect of residual image input on accuracy of the model is analyzed. Results are presented for both, with and without transfer learning to validate the effectiveness of transfer learning.

Large areas of modern power systems are inter-connected for improved power pooling, resulting in higher system inertia. On the other hand, it provides the capability of long-distance transmission of power. Thus, increasing the potential of tie-lines to run at near-maximum capacity. The probability of inter-area oscillations between two areas increases as tie lines are operated closer to full capacity, and this is particularly prevalent near high load density areas. Inter-area stability analysis is becoming more difficult as the current power system becomes more interconnected. This paper implements the energy-sorted Matrix Pencil Method (MPM) on a real-time simulation testbed to perform an online estimation of Low-Frequency Electromechanical Oscillations (LFEOs) present in a power system that is complex and highly interconnected and also implements an oscillation detection method on Phasor Measurement Units (PMUs) data to ensure that energy-sorted MPM is be applied to the relevant data type. As a result, the estimation of modal parameters can be utilized further with confidence and promptly. The proposed process is evaluated on a Klein-Rogers-Kundur test case, and the outcome of the simulation is presented which justifies the effectiveness of energy-sorted MPM in real-time LEFOs.

In this paper a Model Predictive Control (MPC) technique is applied to a direct matrix converter which is used for interfacing of grid to a Wind Energy Conversion System (WECS). Certain LC filters are used along with direct matrix converter to obtain a constant power supply for grid. The paper addresses the work for a WECS with variable speed operation where the primary objective of control is to get constant sinusoidal voltage and constant power and frequency as required to the connected grid. MPC technique is adopted to modulate the matrix converter in order to produces constant and reliable power for grid with different wind speeds for WECS. With MPC, several different control objectives can be easily obtained such as common-mode voltage also the input power factor. The controller is observed for various conditions like variable wind speed, static loads and dynamic loads presented within the grid. The proposed technique for control is reliable, less complex, effective and relevant for application. Simulation study is performed for the proposed method and to verify the capability of the controlled scheme certain results are provided. The MPC technique can be used for micro-grid or smart building with smaller wind conversion systems.

Industrial Control System (ICS) used in critical infrastructure like energy sector, oil and natural gas, chemical processing, waste water treatment etc. are of great importance since it serves to a larger population and hence termed as National Critical Infrastructure (NCI). Security of such infrastructures is essential from physical as well as from cyber point of view. Traditional approach for the security was only physical consisting of creating SOP, access control, physical verification, etc. But due past cyber disturbing events like Stuxnet, Blackenergy-3, Turkey pipeline explosion, Havex, New York dam attack, etc. it is realized that the network security of the connected devices within the infrastructure is also of same importance. Various researches in this field have helped to make the NCI immune to a certain extent as well as resilient against cyber-attacks. One such approach is using Formal Analysis for performing security analysis of the communication protocols used in ICS domain. In this paper a compressive review of all available methods for formal analysis along with CPN modelling is done for analysis of valid as well as invalid states of HART protocol. The Modelling shows various states that the protocol can take during one transaction of communication. Further this analysis can be extended for other Fieldbus protocols for security analysis.

The increasing amount of data on the Internet has been a constant challenge when text cleaning and relevant-text extraction are of interest. One of the areas of focus on the internet is the Dark Web; the data here is much more volatile and dynamic. With more researchers looking for data and extracting information, the algorithms have always been in a state of constant improvement. The solutions currently offered, all work based on text feature extraction algorithms like TF-IDF, Bag of Words, Word2Vec. Discussion on these extraction methods is well documented but a critical evaluation among these algorithms is amiss from standard literature. This paper discusses a balanced approach for tagging extracted data; ELEMENT (Effective Lemmatization, Efficient Management of Extracting Noteworthy Tags) which is a modified form of TF-IDF. Having a balanced approach like ELEMENT will benefit from being able to perform well under any given circumstance. The paper discusses and compares the proposed approach with existing strategies of text feature extraction. This comparison spans across accuracy of feature extraction, efficiency concerning Time and Space, concluding with a simplistic view of the strengths and weaknesses of each algorithm.

This paper proposes a fast and efficient method for magnetic signature reduction of underwater vessels. The magnetic signature reduction is required to protect the ferromagnetic vessels from magnetic anomaly detectors and mines. We propose a novel machine learning-based approach for degaussing of the vessel. This method adds a degree of bias to the evaluated coefficients in order to handle inherent multicollinearity issues. The proposed algorithm is efficient in terms of computational efforts, speed, and accuracy. More than 90% of signature reduction is achieved, assuming that the signature predicted is accurate. The proposed method is validated for a simulated model of prototype submarine.

A wind turbine works under variable load and environmental conditions because of which failure rate has been on the rise. Failure of a gearbox, an integral part of producing wind energy, contributes to 80 % of the total downtime for the wind turbine. For ensuring better utilization of the wind turbines, Fault prognosis and condition monitoring of bearings are of utmost importance as it helps to reduce the downtime by early detection of faults which further increases the power output. In this paper, vibration signals produced and machine learning approach to determine the Remaining Useful Life (RUL) for a degraded bearing is studied. The methodology includes statistical feature extraction analysis with regression models. Further the feature selection is done using Principal Component Analysis (PCA) technique which produces training and testing sets which acts as an input parameter for regression models such as Support Vector Regressor (SVR) and Random Forest (RF). Weibull Hazard Rate Function is used for calculating the RUL of the bearing. Results This study shows the potential application of regression model as an effective tool for degradation performance prediction of bearing.

Obstacle avoidance becomes a very challenging task for an autonomous underwater vehicle (AUV) in an unknown underwater environment during exploration process. Successful control in such case may be achieved using the model-based classical control techniques like PID and MPC but it required an accurate mathematical model of AUV and may fail due to parametric uncertainties, disturbance, or plant model mismatch. On the other hand, model-free reinforcement learning (RL) algorithm can be designed using actual behavior of AUV plant in an unknown environment and the learned control may not get affected by model uncertainties like a classical control approach. Unlike model-based control model-free RL based controller does not require to manually tune controller with the changing environment. A standard RL based one-step Q-learning based control can be utilized for obstacle avoidance but it has tendency to explore all possible actions at given state which may increase number of collision. Hence a modified Q-learning based control approach is proposed to deal with these problems in unknown environment. Furthermore, function approximation is utilized using neural network (NN) to overcome the continuous states and large state-space problems which arise in RL-based controller design. The proposed modified Q-learning algorithm is validated using MATLAB simulations by comparing it with standard Q-learning algorithm for single obstacle avoidance. Also, the same algorithm is utilized to deal with multiple obstacle avoidance problems.

This paper deals with the optimal control of probabilistic Boolean control network (PBCN) in the Markov decision process (MDP) framework that avoids exhaustive search over all inputs. The optimal control problem of PBCN is efficiently tackled if it can be restructured into an equivalent linearly-solvable MDP (LMDP) formulation through the technique of embedding. However, i) the embedded state costs could disagree with original state costs and may even assume negative values, ii) determining the closeness of optimum continuous action of LMDP with discrete actions of the MDP is nontrivial. We propose a novel technique to implement for exact embedding where the embedded state costs match that of the original system. In addition a maximum a posteriori probability based method is proposed to obtain optimum discrete input for PBCN. A couple of examples are presented to illustrate the functioning of the proposed approach.

The researchers established that there exists a portion of the hidden web known as the "Dark Web". The nature of the Dark web intrigued researchers to collect information and present inferences about the information spread and communication taking place on this dark side of the internet.The presented work SpyDark is one such attempt to collect information from the surface and dark web. The user is required to enter the search query, the number of web pages to visit, and specify which network to be accessed (surface or dark web). The crawler traverses all the web pages and extracts information on that page, such as text data, images, hyperlinks, etc. The hyperlinks are stored in the database and later visited by the crawler. The text data is fed to a pre-trained NLP (Natural Language Processing) model, and the output is used to quantify the relevance score. The page is categorized as relevant or irrelevant.

In Industry 4.0, Internet of Thing (IoT) plays a crucial role in solving maintenance problems of a wide range of household as well as industrial equipment. Machine Learning techniques are a vital part of IoT, which deals with an automatic prediction for predictive maintenance systems. ML allows automated fault diagnosis and prognosis for various types of equipment. In this study, Majority Voting Machine Learning approach demonstrated for fault diagnosis of Mechanical components such as rolling element bearing and gearbox with vibration data. Temporal statistical features are extracted from vibration data which are further used as input to ML techniques for training and testing purpose. For analysis, Majority Voting ensemble technique is made by five tuned base classifiers, i.e. Support Vector Machine, k-Nearest Neighbor, Naive Bayes, Multilayer Perceptron and Decision Tree. Results are carried-out for the individual base classifier and Majority Voting technique, and classification performance and accuracy found better with Majority Voting approach than accuracies of individual base classifiers.

Interconnection is the need for a widespread power system for the pooling of power in every crisis, enabling maximum utilization of the grid. Consequently, disturbance on a stressed line can induce low-frequency electromechanical oscillations. Identifying the modes is crucial to take corrective measures and to avoid the potential threat of grid collapse. Matrix Pencil Method is an efficient modal identification method used to identify these oscillations. The basic Matrix Pencil Method identifies trivial modes alongside dominant modes thereby leading to erroneous control. Trivial modes can be distinguished from the dominant modes based on their energy content. Therefore, the paper aims at identifying dominant modes using the Matrix Pencil Method combined with an energy sorting technique. Further, the performance of this approach is comparatively studied with energy sorted Prony Analysis.

Advancements in Artificial Intelligence – oriented computing power and the ever-growing reach of social media have proven to be catalysts in emergence and spread of a new vein of AI generated fake videos known as ’DeepFake’ videos. Such videos are synthesized using generative machine learning models like Generative Adversarial Networks or Variational AutoEncoders and they can achieve high degrees of realism. Spread of sensitive political or obscene content in form of such videos may lead to social distress to the target entity(s). This paper presents a study pertinent to the detection of DeepFake videos using Convolutional Neural Networks (CNNs) with transfer learning. A comparative study of the performance of various models in the detection of tampered videos has been presented. These models are trained (fine-tuned) and tested on a custom dataset encompassing randomly selected labelled frames from videos in the DeepFake Detection Dataset by Google AI and FaceForensics++ dataset.