The growing popularity of Android requires malware detection systems that can keep up with the pace of new software being released. According to a recent study, a new piece of malware appears online every 12 seconds. To address this, we treat Android malware detection as a streaming data problem and explore the use of active online learning as a means of mitigating the problem of labelling applications in a timely and cost-effective manner. Specifically, we develop a semi-supervised active learning framework that incrementally trains online learning models using only samples with low prediction confidence, while detecting concept drift and retraining the models when drift is observed. Our resulting framework achieves accuracies of up to 96% on a balanced dataset, requires as little as 24% of the training data to be labelled, and compensates for concept drift that occurs between the release and labelling of an application. We also consider the broader practicalities of online learning within Android malware detection, and systematically explore the trade-offs between using different static, dynamic and hybrid feature sets to classify malware. We find that features derived from static API calls lead to the best performing models, though models based around lower-dimensional permission and opcode feature sets provide a potentially more practical basis for deployment, with only a marginal deficit in accuracy. Dynamic and hybrid feature sets are found to significantly increase feature extraction costs with no net benefit to predictive performance.

We empirically investigate the effect of class manifold entanglement and the intrinsic and extrinsic dimensionality of the data distribution on the sample complexity of supervised classification with deep ReLU networks. We separate the effect of entanglement and intrinsic dimensionality and show statistically for artificial and real-world image datasets that the intrinsic dimensionality and the entanglement have an interdependent effect on the sample complexity. Low levels of entanglement lead to low increases of the sample complexity when the intrinsic dimensionality is increased, while for high levels of entanglement the impact of the intrinsic dimensionality increases as well. Further, we show that in general the sample complexity is primarily due to the entanglement and only secondarily due to the intrinsic dimensionality of the data distribution.

It is estimated that around 70% of mobile phone users have an Android device. Due to this popularity, the Android operating system attracts a lot of malware attacks. The sensitive nature of data present on smartphones means that it is important to protect against these attacks. Classic signature-based detection techniques fall short when they come up against a large number of users and applications. Machine learning, on the other hand, appears to work well, and also helps in identifying zero-day attacks, since it does not require an existing database of malicious signatures. In this paper, we critically review past works that have used machine learning to detect Android malware. The review covers supervised, unsupervised, deep learning and online learning approaches, and organises them according to whether they use static, dynamic or hybrid features.

Device identification (DI) is one way to secure a network of Internet of Things (IoT) devices, whereby devices identified as suspicious can subsequently be isolated from a network. In this study, we present a machine-learning-based method, IoTDevID, that recognizes devices through the characteristics of their network packets. As a result of using a rigorous feature analysis and selection process, our study offers a generalizable and realistic approach to modeling device behavior, achieving high predictive accuracy across two public data sets. The model’s underlying feature set is shown to be more predictive than existing feature sets used for DI and is shown to generalize to data unseen during the feature selection process. Unlike most existing approaches to IoT DI, IoTDevID is able to detect devices using non-IP and low-energy protocols.

Abstract Parkinson’s disease (PD) is a progressive neurodegenerative disorder that causes abnormal movements and an array of other symptoms. An accurate PD diagnosis can be a challenging task as the signs and symptoms, particularly at an early stage, can be similar to other medical conditions or the physiological changes of normal ageing. This work aims to contribute to the PD diagnosis process by using a convolutional neural network, a type of deep neural network architecture, to differentiate between healthy controls and PD patients. Our approach focuses on discovering deviations in patient’s movements with the use of drawing tasks. In addition, this work explores which of two drawing tasks, wire cube or spiral pentagon, are more effective in the discrimination process. With $$93.5\%$$ 93.5 % accuracy, our convolutional classifier, trained with images of the pentagon drawing task and augmentation techniques, can be used as an objective method to discriminate PD from healthy controls. Our compact model has the potential to be developed into an offline real-time automated single-task diagnostic tool, which can be easily deployed within a clinical setting.

Controlling the dynamics of gene regulatory networks is a challenging problem. In recent years, a number of control methods have been proposed, but most of these approaches do not address the problem of how they could be implemented in practice. In this paper, we consider the idea of using a synthetic regulatory network as a closed-loop controller that can control and respond to the dynamics of a cell's native regulatory network in situ. We explore this idea using a computational model in which both native and synthetic regulatory networks are represented by Boolean networks. We then use an evolutionary algorithm to optimise both the structure and parameters of the synthetic Boolean network. To test this approach, we look at whether controllers can be optimised to target specific steady states in five different Boolean regulatory circuit models. Our results show that in most cases the controllers are able to drive the dynamics of the target system to a specified steady state, often using few interventions, and further experiments using random Boolean networks show that the approach scales well to larger controlled networks.

In this paper, we present a new form of particle swarm optimisation (PSO) in which each particle uses an artificial neural network (ANN) to guide its movements. Information about each of the particle's informants is passed as input to the ANN and the ANN's outputs are then used to select which informant to follow at the next iteration. Using a distributed evolutionary process, each particle's ANN is able to learn about the solution landscape over the course of an optimisation run, potentially allowing the particle to avoid unfavourable regions. An initial evaluation of this approach using a suite of 5 continuous optimisation functions suggests that it improves performance, managing to get consistently closer to the global optima than conventional PSO on all of these problems. An analysis of the trajectories indicates that the behaviour of the algorithm is quite different to conventional PSO, with a much higher degree of exploration than the baseline PSO algorithm.

Neural networks are increasingly being moved to edge computing devices and smart sensors, to reduce latency and save bandwidth. Neural network compression such as quantization is necessary to fit trained neural networks into these resource constrained devices. At the same time, their use in safety-critical applications raises the need to verify properties of neural networks. Adversarial perturbations have potential to be used as an attack mechanism on neural networks, leading to "obviously wrong" misclassification. SMT solvers have been proposed to formally prove robustness guarantees against such adversarial perturbations. We investigate how well these robustness guarantees are preserved when the precision of a neural network is quantized. We also evaluate how effectively adversarial attacks transfer to quantized neural networks. Our results show that quantized neural networks are generally robust relative to their full precision counterpart (98.6%-99.7%), and the transfer of adversarial attacks decreases to as low as 52.05% when the subtlety of perturbation increases. These results show that quantization introduces resilience against transfer of adversarial attacks whilst causing negligible loss of robustness.

This paper describes the artificial epigenetic network, a recurrent connectionist architecture that is able to dynamically modify its topology in order to automatically decompose and solve dynamical problems. The approach is motivated by the behavior of gene regulatory networks, particularly the epigenetic process of chromatin remodeling that leads to topological change and which underlies the differentiation of cells within complex biological organisms. We expected this approach to be useful in situations where there is a need to switch between different dynamical behaviors, and do so in a sensitive and robust manner in the absence of a priori information about problem structure. This hypothesis was tested using a series of dynamical control tasks, each requiring solutions that could express different dynamical behaviors at different stages within the task. In each case, the addition of topological self-modification was shown to improve the performance and robustness of controllers. We believe this is due to the ability of topological changes to stabilize attractors, promoting stability within a dynamical regime while allowing rapid switching between different regimes. Post hoc analysis of the controllers also demonstrated how the partitioning of the networks could provide new insights into problem structure.

Gene regulatory networks are the complex dynamical structures that orchestrate the activities of biological cells. Inappropriate dynamical behaviours, caused by mutations or environmental perturbations, can lead to disease. Control interventions, for example in the form of therapeutic drugs, can lead to recovery from disease. In this paper, we consider how Boolean networks can be used to control a computational model of gene regulatory networks, focusing on the problem of state space targeting in scale-free Boolean networks, an abstract yet realistic model of biological gene regulatory networks. Our results suggest that Boolean networks can be optimised to carry out useful control, and that the approach is relatively scalable. We also take an initial look at the trade-off between the efficacy and efficiency of control, showing that many target networks can be controlled via a relatively small degree of coupling, giving hope that Boolean network controllers could one day be implemented in vivo.

Ensemble classifiers have become a widely researched area in machine learning because they are able to generalise well to unseen data, making them suitable for real world applications. Many approaches implement simple voting techniques, such as majority voting or averaging, to form the overall output. Alternatively, Genetic Algorithms (GAs) can be used to train the ensemble by optimising either binary or floating weights in an average voting scheme to produce evolved linear voting aggregators. Non-linear voting schemes can also be formed by inputting the base classifiers' outputs into a secondary classifier in the form of an expression tree or an Artificial Neural Network; this being subsequently evolved by an Evolutionary Algorithm to optimise the ensemble prediction. This paper aims to firstly, establish the impact of evolving linear aggregators on traditional voting techniques, and secondly whether these linear functions produce more accurate predictions than more complex nonlinear evolved aggregators. The results indicate that optimising the ensemble combination method with GAs offers significant advantages over standard approaches and produces comparable ensembles to those with nonlinear aggregators despite their additional complexity.

Ensemble classifiers have become popular in recent years owing to their ability to produce robust predictive models that generalise well to previously unseen data. In principle, Evolutionary Algorithms (EAs) are well suited to ensemble generation since they result in a pool of trained classifiers. However, in practice they are infrequently used for this purpose. Current research trends in the EA community focus on relatively complex mechanisms for building ensembles, such as co-evolution and multi-objective optimisation. In this paper, we take a back-to-basics approach, studying whether conventional EAs, augmented with simple niching strategies, can be used to form accurate ensembles. We focus on crowding for this, considering both deterministic and probabilistic variants. We also consider the effect of different similarity measures. Our results suggest that simple niching methods can lead to accurate ensemble classifiers and that the choice of similarity measure is not a significant factor. A further study using heterogeneous classifier models within the population showed no added benefit.

Biological organisms exist within environments in which complex nonlinear dynamics are ubiquitous. They are coupled to these environments via their own complex dynamical networks of enzyme-mediated reactions, known as biochemical networks. These networks, in turn, control the growth and behavior of an organism within its environment. In this paper, we consider computational models whose structure and function are motivated by the organization of biochemical networks. We refer to these as artificial biochemical networks and show how they can evolve to control trajectories within three behaviorally diverse complex dynamical systems: 1) the Lorenz system; 2) Chirikov's standard map; and 3) legged robot locomotion. More generally, we consider the notion of evolving dynamical systems to control dynamical systems, and discuss the advantages and disadvantages of using higher order coupling and configurable dynamical modules (in the form of discrete maps) within artificial biochemical networks (ABNs). We find both approaches to be advantageous in certain situations, though we note that the relative tradeoffs between different models of ABN strongly depend on the type of dynamical systems being controlled.

Parkinson's disease is a debilitating neurological condition that affects approximately 1 in 500 people and often leads to severe disability. To improve clinical care, better assessment tools are needed that increase the accuracy of differential diagnosis and disease monitoring. In this paper, we report how we have used evolutionary algorithms to induce classifiers capable of recognizing the movement characteristics of Parkinson's disease patients. These diagnostically relevant patterns of movement are known to occur over multiple time scales. To capture this, we used two different classifier architectures: sliding-window genetic programming classifiers, which model over-represented local patterns that occur within time series data, and artificial biochemical networks, computational dynamical systems that respond to dynamical patterns occurring over longer time scales. Classifiers were trained and validated using movement recordings of 49 patients and 41 age-matched controls collected during a recent clinical study. By combining classifiers with diverse behaviors, we were able to construct classifier ensembles with diagnostic accuracies in the region of 95%, comparable to the accuracies achieved by expert clinicians. Further analysis indicated a number of features of diagnostic relevance, including the differential effect of handedness and the over-representation of certain patterns of acceleration.

A novel bio-inspired architecture comprising three layers is introduced for a six-legged robot in order to generate adaptive rhythmic locomotion patterns using environmental information. Taking inspiration from the intracellular signalling processes that decode environmental information, and considering the emergent behaviours that arise from the interaction of multiple signalling pathways, we develop a decentralised robot controller composed of a collection of artificial signalling networks. Crosstalk, a biological signalling mechanism, is used to couple such networks favouring their interaction. We also apply nonlinear oscillators to model gait generators, which induce symmetric and rhythmical locomotion movements. The trajectories are modulated by a coupled artificial signalling network, which yields adaptive and stable robotic locomotive patterns. Gait trajectories are converted into joint angles by means of inverse kinematics. The architecture is implemented in a simulated version of the real robot T-Hex. Our results demonstrate the ability of the architecture to generate adaptive and periodic gaits.

In this paper, we describe a method for discriminating between thyroid cell lines. Five commercial thyroid cell lines were obtained, ranging from non-cancerous to cancerous varieties. Raman spectroscopy was used to interrogate native cell biochemistry. Following suitable normalisation of the data, implicit context representation Cartesian genetic programming was then used to search for classifiers capable of distinguishing between the spectral fingerprints of the different cell lines. The results are promising, producing comprehensible classifiers whose output values correlate with biological aggressiveness.

In this paper, a revised form of implicit context representation Cartesian genetic programming is used in the development of a diagnostic tool for the assessment of patients with neurological dysfunction such as Alzheimer's disease. Specifically, visuo-spatial ability is assessed by analysing subjects' digitised responses to a simple figure copying task using a conventional test environment. The algorithm was trained to distinguish between classes of visuo-spatial ability based on responses to the figure copying test by 7-11 year old children in which visuo-spatial ability is at varying stages of maturity. Results from receiver operating characteristic (ROC) analysis are presented for the training and subsequent testing of the algorithm and demonstrate this technique has the potential to form the basis of an objective assessment of visuo-spatial ability.

In previous work, we have shown how an evolutionary algorithm with a clustered population can be used to concurrently discover multiple regulatory motifs present within the promoter sequences of co-expressed genes. In this paper, we extend the algorithm by co-evolving a population of Boolean classification rules in parallel with the motif population. Results using synthetic data suggest that this approach allows poorly conserved motifs to be identified in promoter sequences an order of magnitude longer than using population clustering alone, whilst results using muscle-specific promoter data show the algorithm is able to evolve meaningful sequence classifiers in parallel with motifs-suggesting that co-evolution provides a suitable framework for composite motif discovery within eukaryotic sequences.

This paper describes a novel evolutionary algorithm for regulatory motif discovery in DNA promoter sequences. The algorithm uses data clustering to logically distribute the evolving population across the search space. Mating then takes place within local regions of the population, promoting overall solution diversity and encouraging discovery of multiple solutions. Experiments using synthetic data sets have demonstrated the algorithm's capacity to find position frequency matrix models of known regulatory motifs in relatively long promoter sequences. These experiments have also shown the algorithm's ability to maintain diversity during search and discover multiple motifs within a single population. The utility of the algorithm for discovering motifs in real biological data is demonstrated by its ability to find meaningful motifs within muscle-specific regulatory sequences.

The functionality model is a new approach in enzyme genetic programming which enables the evolution of variable length solutions whilst preserving local context. This paper introduces the model and presents an analysis of crossover and the evolution of program size.

The work reported in the paper follows from the hypothesis that better performance in certain domains of artificial evolution can be achieved by adhering more closely to the features that make natural evolution effective within biological systems. An important issue in evolutionary computation is the choice of solution representation. Genetic programming, whilst borrowing from biology in the evolutionary axis of behaviour, remains firmly rooted in the artificial domain with its use of a parse tree representation. Following concerns that this approach does not encourage solution evolvability, the paper presents an alternative method modelled upon representations used by biology. Early results are encouraging, demonstrating that the method is competitive when applied to problems in the area of combinatorial circuit design.