Attacks on computer networks have increased significantly in recent days, due in part to the availability of sophisticated tools for launching such attacks as well as the thriving underground cyber-crime economy to support it. Over the past several years, researchers in academia and industry used machine learning (ML) techniques to design and implement Intrusion Detection Systems (IDSes) for computer networks. Many of these researchers used datasets collected by various organizations to train ML classifiers for detecting intrusions. In many of the datasets used in training ML classifiers in such systems, data are imbalanced (i.e., not all classes had equal number of samples). ML classifiers trained with such imbalanced datasets may produce unsatisfactory results. Traditionally, researchers used over-sampling and under-sampling for balancing data in datasets to overcome this problem. In this work, in addition to random over-sampling, we also used a synthetic data generation method, called Conditional Generative Adversarial Network (CTGAN), to balance data and study their effect on the performance of various widely used ML classifiers. To the best of our knowledge, no one else has used CTGAN to generate synthetic samples to balance intrusion detection datasets. Based on extensive experiments using widely used datasets NSL-KDD and UNSW-NB15, we found that training ML classifiers on datasets balanced with synthetic samples generated by CTGAN increased their prediction accuracy by up to 8% and improved their MCC score by up to 13%, compared to training the same ML classifiers over imbalanced datasets. We also show that this approach consistently performs better than some of the recently proposed state-of-the-art IDSes on both datasets. Our experiments also demonstrate that the accuracy of some ML classifiers trained over datasets balanced with random over-sampling decline compared to the same ML classifiers trained over original imbalanced dataset.

Despite employing the architectures designed for high service reliability and availability, cloud computing systems do experience service outages and performance slowdown. In addition to these, large-scale cloud systems experience failures in their hardware and software components which often result in node and application (e.g., jobs and tasks) failures. Therefore, to build a reliable cloud system, it is important to understand and characterize the observed failures. The goal of this work is to identify the key features that correlate to application failures in cloud and present a failure prediction model that can correctly predict the outcome of a task or job before it actually finishes, fails or gets killed. To accomplish this, we perform a failure characterization study of the Google cluster workload trace. Our analysis reveals that, there is a significant consumption of resources due to failed and killed jobs. We further explore the potential for failure prediction in cloud applications so that we can reduce the wastage of resources by better managing the jobs and tasks that ultimately fail or get killed. For this, we propose a prediction method based on a special type of Recurrent NeuralNetwork (RNN) named Long Short-Term Memory Network(LSTM) to identify application failures in cloud. It takes resource usage measurements or performance data for each job and task, and the goal is to predict the termination status (e.g., failed and finished etc.) of them. Our algorithm can predict task failures with 87%accuracy and achieves a true positive rate of 85% and false positive rate of 11%.