Machine Learning and Metaheuristics Algorithms, and Applications

Facial Action Units (AUs) correspond to the deformation/contraction of individual or combinations of facial muscles. As such, each AU affects just a small portion of the face, with deformations that are asymmetric in many cases. Analysing AUs in 3D is particularly relevant for the potential applications it can enable. In this paper, we propose a solution for AUs detection by developing on a newly defined 3D Morphable Model (3DMM) of the face. Differently from most of the 3DMMs existing in the literature, that mainly model global variations of the face and show limitations in adapting to local and asymmetric deformations, the proposed solution is specifically devised to cope with such difficult morphings. During a learning phase, the deformation coefficients are learned that enable the 3DMM to deform to 3D target scans showing neutral and facial expression of a same individual, thus decoupling expression from identity deformations. Then, such deformation coefficients are used to train an AU classifier. We experimented the proposed approach in a difficult cross-dataset experiment, where the 3DMM is constructed on one dataset (BU-3DFE) and tested on a different one (Bosphorus). Results evidence that effective AU detection is obtained by SVM learning of deformation coefficients from a small training set.

The paper proposes to design and develop a smart authentication system in laboratory as a part of security and surveillance. To address the unauthorized entry in the laboratory, a smart alert system is designed and developed. The authentic entry to any laboratory will reduce the student response to hazards and accidents, risks to acceptable levels. The proposed methodology uses face detection and recognition techniques for the student authentication. Based on the results, the attendance is updated in the attendance data base if the authorized users enter the laboratory else the details will be sent to the course instructors through the registered mails. The authentic student is also verified for wearing the personal protective equipment during the entry to the laboratory. By this, we can reduce the vandalism occurring in laboratories and maintain the integrity.

Iris recognition technologies applied to produce comprehensive and correct biometric identification of people in numerous large-scale data of humans. Additionally, the iris is stable over time, i.e., iris biometric knowledge offers links between biometric characteristics and people. The e-business and e-governance require more machine-driven iris recognition. It has millions of iris images that are in near-infrared illumination. It is used for people’s identity. A variety of applications for surveillance and e-business will embody iris pictures that are unit non-heritable below visible illumination. The self-learned iris features are created by the convolution neural network (CNN), give more accuracy than handcrafted feature iris recognition. In this paper, a modified iris recognition system is introduced using deep learning techniques along with multi-class SVM for matching. We use the Poly-U database, which is from 209 subjects. CNN with softmax cross-entropy loss gives the most accurate matching of testing images. This method gives better results in terms of EER. We analyzed the proposed architecture on other publicly available databases through various experiments.

The proposed system, Gaze Fusion - Deep Neural Network Model (GFDM) has utilized transfer learning approach to discriminate subject’s eye tracking data in the form of fusion map into two classes: glaucoma and normal. We have fed eye tracking data in the form of fusion maps of different participants to Deep Neural Network (DNN) model which is pretrained with ImageNet weights. The experimental results of the GFDM show that fusion map dissimilar to pretrained model’s dataset can give better understanding of glaucoma. The model also show the part of the screen where participants has the difficulty in viewing. GFDM has compared with traditional machine learning models such as Support Vector Classifier, Decision Tree classifier and ensemble classifier and shown that the proposed model outperforms other classifiers. The model has Area Under ROC Curve (AUC) score 0.75. The average sensitivity of correctly identifying glaucoma patients is 100% with specificity value 83%.

This paper presents a deep learning based solution for identification of normal and abnormal candle flames, controlled and uncontrolled flames. Candle flames affected by external factors like wind, improper combustion of fuel etc. Proposed CNN based deep neural network can successfully classify the stable and unstable candle flame with an accuracy of 67% for generated test set and an accuracy of 83% for random images taken from open source on internet.

The research on automatic emotional recognition has been increased drastically because of its significant influence on various applications such as treatment of the illness, educational practices, decision making, and the development of commercial applications. Using Machine Learning (ML) models, we have been trying to determine the emotion accurately and precisely from the facial expressions. But it requires a colossal number of resources in terms of data as well as computational power and can be time-consuming during its training. To solve these complications, meta-learning has been introduced to train a model on a variety of learning tasks, which assists the model to generalize the novel learning tasks using a restricted amount of data. In this paper, we have applied one of the meta-learning techniques and proposed a model called MLARE(Meta Learning Approach to Recognize Emotions) that recognizes emotions using our in-house developed dataset AED-2 (Amrita Emotion Dataset-2) which has 56 images of subjects expressing seven basic emotions viz., disgust, sad, fear, happy, neutral, anger, and surprise. It involves the implementation of the Siamese network which estimates the similarity between the inputs. We could achieve 90.6% of overall average accuracy in recognizing emotions with the state-of-the-art method of one-shot learning tasks using the convolutional neural network in the Siamese network.

A model which predicts the activities of each individual in a household is developed. This model is used to simulate the activities of 25,000 households in a town in Kerala (a state in southern region of India). A fuzzy-logic based approach is used to estimate the probabilities of an individual to be in a particular state/activity. Then an optimization problem is formulated to compute the activity transitions. Further, the activity transitions of the individuals within a house are tied in an appropriate way. These activity transitions are then used to simulate a Markov chain of the activities for a sample set of households in a town in Kerala.

Prediction of stock prices has been an important area of research for a long time. While supporters of the efficient market hypothesis believe that it is impossible to predict stock prices accurately, there are formal propositions demonstrating that accurate modeling and designing of appropriate variables may lead to models using which stock prices and stock price movement patterns can be very accurately predicted. Researchers have also worked on technical analysis of stocks with a goal of identifying patterns in the stock price movements using advanced data mining techniques. In this work, we propose an approach of hybrid modeling for stock price prediction building different machine learning and deep learning-based models. For the purpose of our study, we have used NIFTY 50 index values of the National Stock Exchange (NSE) of India, during the period December 29, 2014 till July 31, 2020. We have built eight regression models using the training data that consisted of NIFTY 50 index records from December 29, 2014 till December 28, 2018. Using these regression models, we predicted the open values of NIFTY 50 for the period December 31, 2018 till July 31, 2020. We, then, augment the predictive power of our forecasting framework by building four deep learning-based regression models using long-and short-term memory (LSTM) networks with a novel approach of walk-forward validation. Using the grid-searching technique, the hyperparameters of the LSTM models are optimized so that it is ensured that validation losses stabilize with the increasing number of epochs, and the convergence of the validation accuracy is achieved. We exploit the power of LSTM regression models in forecasting the future NIFTY 50 open values using four different models that differ in their architecture and in the structure of their input data. Extensive results are presented on various metrics for all the regression models. The results clearly indicate that the LSTM-based univariate model that uses one-week prior data as input for predicting the next week’s open value of the NIFTY 50 time series is the most accurate model.

Stock market prediction is a tool to maximize the investor’s money and minimize the risk associated with it. This paper proposes a new machine learning based model to predict the stock market price. The proposed model is an improved version of existing Salp Swarm Optimizer (SSO) which is integrated with Least Square-Support Vector Machine (LSSVM). The improved version is a hybrid meta-heuristics algorithm, which is a combination of SSO and Adaptive \(\upbeta \)-Hill Climbing (A\(\upbeta \)-HC) algorithm. The proposed model selects the best hyperparameters for LSSVM to avoid over fitting and local minima, and in turn increases the model’s accuracy. It is evaluated on four standard and publicly available stock market datasets, and its result is compared with some popular meta-heuristic algorithms. The results show that our proposed model performs better than the existing models in most of the cases. The source code for our proposed algorithm is available in https://​github.​com/​singhaviseq/​SSA-ABHC.

The indigenous part of all living organisms in the world is food. As the world population increases, the production and consumption of food also increases. Since the population progresses in a rapid manner, the productivity of the food materials may not be sufficient for feeding all the people in the world. There rises the cause of food adulteration and food fraud. Adulteration is the process of adding a foreign substance to the food material which affects the natural quality of the food. As the amount of adulterants increases, the toxicity also increases. Machine learning techniques has been used previously to automate the prediction of food adulteration under normal scenarios. In this paper, we use different machine learning technique for finding food adulteration from milk data sets. This paper surveys the different concepts used in automating the detection of food adulteration and discusses the experimental results obtained by applying machine learning algorithms like Naive Bayes, Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Artificial Neural networks (ANN), Linear Regression, and Ensemble methods. The accuracy of the models ranged from 79% to 89%. Ensemble method outperformed other algorithms with an accuracy of 89% and Linear Regression showed least accuracy of 79%. Artificial Neural networks showed an accuracy of almost 87%. SVM and Naïve Bayes showed accuracy 84% and 80% respectively.

Despite unceasing debate about it’s pros and cons, exams and standardized testing have emerged as the main mode of evaluation and comparison in our increasingly competitive world. Inevitably, some examinees attempt to illegally gain an unfair advantage over other candidates by indulging in cheating and malpractice. Even a single case of examination malpractice can destroy an Examination body’s credibility and even lead to costly and time-consuming legal proceedings.

Our paper attempts to strategically allot examinees in specific seats and rooms, such as to mitigate the overall probability of malpractice. It involves examining multiple crucial factors such as subject similarity, distancing between examinees, and human field of vision to find the most optimal seating arrangement. We have exploited the property of Evolutionary Genetic Algorithms to find globally optimal or close to optimal solutions in an efficient time for this otherwise NP-complete permutation problem.

With the emergence of big data and business analytics, many new concepts and business models driven by data have been introduced in recent years. Does “big” really matter? We attempt to explain when and why big does not matter in many business cases. In those occasions where “Big” does matter, we outline the data strategy framework that differentiates the degree of big data requirements. In conclusion, we offer practical advice on strategic usage of big data for best practice management and sound decision-making.

Electricity energy billing system is prevalent in most of the places in the world. Also digitization of these electricity bills has also been successfully implemented in various underdeveloped countries as well. The vast amount of data is available regarding the energy consumption of consumers. In this paper we consider a case study of one city, about which we have electricity energy data for several years. We first classify consumers based on their average energy usage via clustering algorithms. We also have survey data of several houses. In that survey, we have building information, family information and also appliance information. We use various regression techniques to disaggregate the energy usage corresponding to various appliances.

In the family of deep neural network models, deeper the model is, the longer it takes to predict and larger the memory space it utilizes. It is very much likely that use-cases have constraints to be respected, especially on embedded devices, i.e, low powered, memory-constrained systems. Finding a suitable model under constraints is repeated trial-and-error to find optimal trade-off. A novel technique known as Neural Ordinary Differential Equation Networks (ODENet) was proposed in NeurIPS2018, where instead of a distinct arrangement of internal hidden layers of a Residual Neural Network (RNN), they used parametrized derivatives of internal states in the neural system. Any differential equation solver can be used to calculate the final output. These models have constant depth and can trade between speed and accuracy. We propose a methodology for Traffic Sign Detection using ODENet and subsequently conclude that ODENets are more robust and perform better in comparison to ResNets. We also conclude that though training time is high in ODENets, they can trade-off between speed and accuracy when it comes to both training and testing.

Benchmark datasets are the inevitable tool required to scrutinize vulnerabilities and tools in network security. Current datasets lack correlation between normal and the real-time network traffic. Behind every evaluation and establishment of attack detection, such datasets are the cornerstone deployed by research community. Creating our own dataset is a herculean task. Hence analyzing the subsisting datasets aids to provide a thorough clarity on the effectiveness when deployed in real time environments. This paper work focus on analysis and comparison of UNSW-NB15 with NSL-KDD dataset based on performance analysis and accuracy using machine learning classifiers. Feasibility, reliability and dependability of the dataset is reviewed and discussed by considering various performance measures such as precision, recall, F-score, specificity using various machine learning classifiers Naïve Bayes, Logistic Regression, SMO, J48 and Random Forest. Experimental results give out its noticeable classification accuracy of 0.99 with the random forest classifier having 0.998 recall and specificity 0.999 respectively. Research studies reveal the fact that threat diagnosis using conventional dataset and sophisticated technologies cover only 25% of threat taxonomy and hence the poor performance of existing intrusion detection systems. Thorough analysis and exploration of the dataset will pave the way for the outstanding performance of the intelligent Intrusion Detection System.

When machine learning models are built with non-stationary data their performance will naturally decrease over time due to concept drift, shifts in the underlying distribution of the data. A common solution is to retrain the machine learning model which can be expensive, both in obtaining new labeled data and in compute time. Traditionally many approaches to concept drift detection operate upon streaming data. However drift is also prevalent in semi-stationary data such as web data, social media, and any data set which is generated from human behaviors. Changing web technology causes concept drift in the website data that is used by phishing detection models. In this work, we create “Autoencoder Drift Detection” (ADD) an unsupervised approach for a drift detection mechanism that is suitable for semi-stationary data. We use the reconstruction error of the autoencoder as a proxy to detect concept drift. We use ADD to detect drift in a phishing detection data set which contains drift as it was collected over one year. We also show that ADD is competitive within ±24% with popular streaming drift detection algorithms on benchmark drift datasets. The average accuracy on the phishing data set is .473 without drift detection and using ADD is increased to .648.

Obfuscation techniques are used by malware authors to conceal malicious code and surpass the antivirus scanning. Machine Learning techniques especially deep learning techniques are strong enough to identify obfuscated malware samples. Performance of deep learning model on obfuscated malware detection is compared with conventional machine learning models like Random Forest (RF), Classification and Regression Trees (CART) and K Nearest Neighbour (KNN). Both Static (hardware and permission) and dynamic features (system calls) are considered for evaluating the performance. The models are evaluated using metrics which are precision, recall, F1-score and accuracy. Obfuscation transformation attribution is also addressed in this work using association rule mining. Random forest produced best outcome with F1-Score of 0.99 with benign samples, 0.95 with malware and 0.94 with obfuscated malware with system calls as features. Deep learning network with feed forward architecture is capable of identifying benign, malware, obfuscated malware samples with F1-Score of 0.99, 0.96 and 0.97 respectively.

With the rapid growth of the Internet and smartphone and wireless communication-based applications, new threats, vulnerabilities, and attacks also increased. The attackers always use communication channels to violate security features. The fast-growing of security attacks and malicious activities create a lot of damage to society. The network administrators and intrusion detection systems (IDS) were also unable to identify the possibility of network attacks. However, many security mechanisms and tools are evolved to detect the vulnerabilities and risks involved in wireless communication. Apart from that machine learning classifiers (MLCs) also practical approaches to detect intrusion attacks. These MLCs differentiated the network traffic data as two parts one is abnormal and other regular. Many existing systems work on the in-depth analysis of specific attacks in network intrusion detection systems. This paper presents a comprehensive and detailed inspection of some existing MLCs for identifying the intrusions in the wireless network traffic. Notably, we analyze the MLCs in terms of various dimensions like feature selection and ensemble techniques to identify intrusion detection. Finally, we evaluated MLCs using the “NSL-KDD” dataset and summarize their effectiveness using a detailed experimental evolution.