As new cyberattacks are launched against systems and networks on a daily basis, the ability for network intrusion detection systems to operate efficiently in the big data era has become critically important, particularly as more low-power Internet-of-Things (IoT) devices enter the market. This has motivated research in applying machine learning algorithms that can operate on streams of data, trained online or “live” on only a small amount of data kept in memory at a time, as opposed to the more classical approaches that are trained solely offline on all of the data at once. In this context, one important concept from machine learning for improving detection performance is the idea of “ensembles”, where a collection of machine learning algorithms are combined to compensate for their individual limitations and produce an overall superior algorithm. Unfortunately, existing research lacks proper performance comparison between homogeneous and heterogeneous online ensembles. Hence, this paper investigates several homogeneous and heterogeneous ensembles, proposes three novel online heterogeneous ensembles for intrusion detection, and compares their performance accuracy, run-time complexity, and response to concept drifts. Out of the proposed novel online ensembles, the heterogeneous ensemble consisting of an adaptive random forest of Hoeffding Trees combined with a Hoeffding Adaptive Tree performed the best, by dealing with concept drift in the most effective way. While this scheme is less accurate than a larger size adaptive random forest, it offered a marginally better run-time, which is beneficial for online training.

The prevalence of Internet-of-Things (IoT) technologies and the ubiquity of networked sensors and actuators in many industrial control systems (ICS) have led to the exposure of critical infrastructure in our society to malicious activities and cyber threats. Programmable logic controllers (PLCs) are embedded devices that automate ICS processes. PLCs, which serve as the heart of ICS, are vulnerable to attacks and system malfunctions like other embedded devices. Because PLCs are widely used to control ICS physical processes, attacks against PLCs can cause irreparable damage to enterprises and even loss of life. However, due to the unique and proprietary architecture of PLCs, it is not easy to apply traditional tools and techniques for PLC protection. This work presents an unsupervised learning approach for anomaly detection in ICS based on neural networks with one class objective function and additional regularization term. This technique combines the abilities of neural networks to learn complex relationships with a one-class objective function and a regularization term for separating normal conditions from anomalous operations. The newly introduced regularization term provides a model-tuning mechanism based on specific industrial requirements and performance metrics of interest (i.e., precision or recall). The model was evaluated on a recent real-world ICS dataset: the Secure Water Treatment (SWaT) dataset. The proposed technique’s performance is compared with previous work, showing improvements in terms of scalability and attack detection capability, proving that the proposed technique is suitable for use in real ICS scenarios. The proposed method with the regularization term demonstrated superior recall values in 15 out of the 36 attack scenarios in the SWaT dataset, which is the largest of any published methods in the literature. A qualitative analysis of the proposed technique on the SWaT security showdown event data further proves the technique’s high anomaly detection ability on real-time injected attacks.

Precision medicine informatics is a field of research that incorporates learning systems that generate new knowledge to improve individualized treatments using integrated data sets and models. Given the ever-increasing volumes of data that are relevant to patient care, artificial intelligence (AI) pipelines need to be a central component of such research to speed discovery. Applying AI methodology to complex multidisciplinary information retrieval can support efforts to discover bridging concepts within collaborating communities. This dovetails with precision medicine research, given the information rich multi-omic data that are used in precision medicine analysis pipelines. In this perspective article we define a prototype AI pipeline to facilitate discovering research connections between bioinformatics and clinical researchers. We propose building knowledge representations that are iteratively improved through AI and human-informed learning feedback loops supported through crowdsourcing. To illustrate this, we will explore the specific use case of nonalcoholic fatty liver disease, a growing health care problem. We will examine AI pipeline construction and utilization in relation to bench-to-bedside bridging concepts with interconnecting knowledge representations applicable to bioinformatics researchers and clinicians.

Counterfactual explanations are gaining in popularity as a way of explaining machine learning models. Counterfactual examples are generally created to help interpret the decision of a model. In that case, if a model makes a certain decision for an instance, the counterfactual examples of that instance reverse the decision of the model. Counterfactual examples can be created by craftily changing particular feature values of the instance. Though counterfactual examples are generated to explain the decision of machine learning models, we have already explored that counterfactual examples can be used for effective data augmentation. In this work, we want to explore what kind of counterfactual examples work best for data augmentation. In particular, we want to generate counterfactual examples from two perspectives: proximity and diversity. We want to observewhich perspective works best in this regard. We demonstrate the efficacy of these approaches on the widely used “Adult-Income” dataset. We consider several scenarios where we do not have enough data and use each of these approaches to augment the dataset. We compare these two approaches and discuss the implications of the results.

Migratory waterfowl (i.e., ducks, geese, and swans) management relies on landscape bioenergetic models to inform on-the-ground habitat conditions and conservation practices. Therefore, conservation planners rely on accurate predictions of wetland habitats for waterfowl at regional scales. Unharvested flooded corn is a popular management tool on public and private lands that greatly increases landscape-level energy compared to other wetlands; thus, landscape bioenergetic models are particularly sensitive to these habitat features. Despite their importance to conservation planning and implementation, the abundance and distribution of unharvested flooded corn fields across North America is unknown. Furthermore, training data is difficult to collect and accurate predictions are challenging given their unique attributes and discreteness at landscape-level lens. Advances in multispectral imagery and deep learning algorithms may enable continuous and autonomous detection of these habitat features. Therefore, we conducted modeling experiments using training data of unharvested flooded corn fields in West Tennessee and multispectral imagery collected from Sentinel-2 satellite missions. We performed several experiments using individual band combination composites and/or vegetation indices to identify optimal bands using MRUNET architectures. We subsequently used 3 ensemble models of important individual networks. We found the use of multispectral bands was necessary and although the CIR composite and OSAVI index improved precision, the 12-band composite increased recall, the metric we were most interested in. Moreover, all ensembles exhibited poor performance. Here, we present results of our initial modeling experiments and suggest future modeling exercises including temporal image and vegetation index stacking using multi-modal and/or recurrent neural network architectures.

Model compression is increasingly popular in the domain of deep learning. When addressing practical problems that use complex neural network models, the availability of computational resources can pose a significant challenge. While smaller models may provide more efficient solutions, they often come at the cost of accuracy. To tackle this problem, researchers often use model compression techniques to transform large, complex models into simpler, faster models. These techniques aim to reduce the computational cost while minimizing the loss of accuracy. The majority of the model compression research focuses exclusively on model accuracy and size/speedup as performance metrics. This paper explores how different methods of model compression impact the fairness/bias of a model. We conducted our experiments using the COMPAS Recidivism Racial Bias dataset. We evaluated a variety of model compression techniques across multiple bias groups. Our findings indicate that the type and amount of compression have substantial impact on both the accuracy and fairness/bias of the model.