The emerging technology of Lab-on-Chips has made a profound impact in the field of healthcare, biochemistry, and molecular biology involving key tasks such as clinical trials, drug therapy, DNA-sequencing, among other tasks. In particular, digital microfluidic biochips have found versatile applications because of the simplicity of operations and low operational cost. Unfortunately, the ease of programmability and controllability in these biochips open doors to severe infringement of privacy and security, which in turn jeopardizes the trustworthiness of bioprotocols. The insertion of Trojans in these biochips may result in deadly outcomes and thus, ensuring the security of biochips has become a major challenge so as to prevent malicious alterations, data theft, cybercrime, sabotage, breach of confidence, and vandalism. Conventional techniques for Trojan detection such as side-channel analysis, runtime scan, and reverse engineering suffer from many shortcomings that compromise time, cost, and the reliability of bioprotocols. In this paper, we present new countermeasures to protect digital microfluidic biochips against Trojan attacks based on image classification of the running bioprotocol. A deep-learning approach is employed to classify the observed snapshot sequences as good or infected. Our results on five real-life bioprotocols demonstrate that the proposed classifier achieves 98.52% accuracy while preserving the security and functionalities of the digital microfluidic biochip.

Advancements in software defect prediction (SDP) to handle the scenario of no or limited historical data have introduced the concept of cross-project defect prediction (CPDP). CPDP using machine learning (ML) algorithms has been the staple research area for all software practitioners in the SDP domain. An important assumption in ML algorithms is that both train and test data must follow similar data distribution for better accuracy. These assumptions may hold in the within-project defect prediction (WPDP) scenario where both train and test data belong to the same project. However, it is impossible in the CPDP scenario where the train and test data belong to different projects. So, in the CPDP scenario, researchers tried to use a matched metrics approach to handling this issue. However, in this case, there may be an issue if only a small-sized source (train) dataset matches the data distribution with the target (test) dataset, leading to an insufficient training dataset. Hence, we have proposed a cross-project data preprocessing method, namely knowledge transfer from target data to source data using correlation (KTTSC), to handle this issue and hence to improve the CPDP accuracy of ML models. The experimental results demonstrate that using the dropout regularization-based deep neural network, k nearest neighbor, decision tree, logistic regression, and Naive Bayes classifiers with the proposed KTTSC method show an improvement of 22%, 17%, 23.2%, 13.5%, and 9.5%, respectively, in terms of average AUC scores as compared to the traditional CPDP method and an improvement in the range of 6.6% to 11.1% as compared to existing works on CPDP.

Heterogeneous defect prediction (HDP) is a promising research area in the software defect prediction domain to handle the unavailability of the past homogeneous data. In HDP, the prediction is performed using source dataset in which the independent features (metrics) are entirely different than the independent features of target dataset. One important assumption in machine learning is that independent features of the source and target datasets should be relevant to each other for better prediction accuracy. However, these assumptions do not generally hold in HDP. Further in HDP, the selected source dataset for a given target dataset may be of small size causing insufficient training. To resolve these issues, we have proposed a novel heterogeneous data preprocessing method, namely, Transfer of Data from Target dataset to Source dataset selected using Relevance score (TDTSR), for heterogeneous defect prediction. In the proposed approach, we have used chi-square test to select the relevant metrics between source and target datasets and have performed experiments using proposed approach with various machine learning algorithms. Our proposed method shows an improvement of at least 14% in terms of AUC score in the HDP scenario compared to the existing state of the art models.

Imbalanced data is a significant issue in software fault prediction. It is very challenging for software engineers to handle imbalanced software fault data for the early prediction of software faults. In the last two decades, many researchers have used synthetic minority oversampling technique (SMOTE), SMOTE for regression and other such techniques to preprocess the imbalanced software fault data. However, these preprocessing techniques do not produce consistently good accuracy, especially in inter release, and cross project fault prediction. The learning of imbalanced fault data for prediction of the number of software faults has not been explored in depth so far. To deal with this scenario, we have explored an efficient machine learning technique, namely extreme learning machine (ELM) for prediction of the number of software faults. Furthermore, a new variant of ELM, namely weighted regularization ELM, is proposed to generalize the imbalanced data to balanced data. To validate the proposed imbalanced learning model, we have used 26 open source PROMISE software fault datasets and three prediction scenarios, intra release, inter release, and cross project. We have conducted the experiments for prediction of the number of faults. The experimental results showed that the proposed approach led to improved performance.

Effort Estimation is a very challenging task in the software development life cycle. Inaccurate estimations may cause client dissatisfaction and thereby, decrease the quality of the product. Considering the problem of software cost and effort estimation, it is conceivable to call attention to that the estimation procedure considers the qualities present in the data set, as well as the aspects of the environment in which the model is embedded. Existing literature have the instances where machine learning techniques have been used to estimate the effort required to develop any software. Yet it is quite uncertain for any particular model to perform well with all the data sets. In this paper, Multi-Layer Perceptron (MLPNN) and its ensembles are explored in order to improve the performance of software effort estimation process. Firstly, MLPNN, Ridge-MLPNN, Lasso-MLPNN, Bagging-MLPNN, and AdaBoost-MLPNN models are developed and, then, the performance of these models are compared on the basis of R2score to find the best model fitting this dataset. Results obtained from the study demonstrate that the R2score of AdaBoost-MLPNN is 82.213%, which is highest among all the models.