Detection of cyber attacks against vehicles is of growing interest. As vehicles typically afford limited processing resources, proposed solutions are rule-based or lightweight machine learning techniques. We argue that this limitation can be lifted with computational offloading commonly used for resource-constrained mobile devices. The increased processing resources available in this manner allow access to more advanced techniques. Using as case study a small four-wheel robotic land vehicle, we demonstrate the practicality and benefits of offloading the continuous task of intrusion detection that is based on deep learning. This approach achieves high accuracy much more consistently than with standard machine learning techniques and is not limited to a single type of attack or the in-vehicle CAN bus as previous work. As input, it uses data captured in realtime that relate to both cyber and physical processes, which it feeds as time series data to a neural network architecture. We use both a deep multilayer perceptron and a recurrent neural network architecture, with the latter benefitting from a long-short term memory hidden layer, which proves very useful for learning the temporal context of different attacks. We employ denial of service, command injection and malware as examples of cyber attacks that are meaningful for a robotic vehicle. The practicality of computation offloading depends on the resources afforded onboard and remotely, and the reliability of the communication means between them. Using detection latency as the criterion, we have developed a mathematical model to determine when computation offloading is beneficial given parameters related to the operation of the network and the processing demands of the deep learning model. The more reliable the network and the greater the processing demands, the greater the reduction in detection latency achieved through offloading.

The modern Internet of Things (IoT)-based smart home is a challenging environment to secure: devices change, new vulnerabilities are discovered and often remain unpatched, and different users interact with their devices differently and have different cyber risk attitudes. A security breach’s impact is not limited to cyberspace, as it can also affect or be facilitated in physical space, for example, via voice. In this environment, intrusion detection cannot rely solely on static models that remain the same over time and are the same for all users. We present MAGPIE, the first smart home intrusion detection system that is able to autonomously adjust the decision function of its underlying anomaly classification models to a smart home’s changing conditions (e.g., new devices, new automation rules and user interaction with them). The method achieves this goal by applying a novel probabilistic cluster-based reward mechanism to non-stationary multi-armed bandit reinforcement learning. MAGPIE rewards the sets of hyperparameters of its underlying isolation forest unsupervised anomaly classifiers based on the cluster silhouette scores of their output. Experimental evaluation in a real household shows that MAGPIE exhibits high accuracy because of two further innovations: it takes into account both cyber and physical sources of data; and it detects human presence to utilise models that exhibit the highest accuracy in each case. MAGPIE is available in open-source format, together with its evaluation datasets, so it can benefit from future advances in unsupervised and reinforcement learning and be able to be enriched with further sources of data as smart home environments and attacks evolve.

Semantic social engineering attacks are a pervasive threat to computer and communication systems. By employing deception rather than by exploiting technical vulnerabilities, spear-phishing, obfuscated URLs, drive-by downloads, spoofed websites, scareware, and other attacks are able to circumvent traditional technical security controls and target the user directly. Our aim is to explore the feasibility of predicting user susceptibility to deception-based attacks through attributes that can be measured, preferably in real-time and in an automated manner. Toward this goal, we have conducted two experiments, the first on 4333 users recruited on the Internet, allowing us to identify useful high-level features through association rule mining, and the second on a smaller group of 315 users, allowing us to study these features in more detail. In both experiments, participants were presented with attack and non-attack exhibits and were tested in terms of their ability to distinguish between the two. Using the data collected, we have determined practical predictors of users' susceptibility against semantic attacks to produce and evaluate a logistic regression and a random forest prediction model, with the accuracy rates of. 68 and. 71, respectively. We have observed that security training makes a noticeable difference in a user's ability to detect deception attempts, with one of the most important features being the time since last self-study, while formal security education through lectures appears to be much less useful as a predictor. Other important features were computer literacy, familiarity, and frequency of access to a specific platform. Depending on an organisation's preferences, the models learned can be configured to minimise false positives or false negatives or maximise accuracy, based on a probability threshold. For both models, a threshold choice of 0.55 would keep both false positives and false negatives below 0.2.

Industrial cyber-physical systems (ICPS) are widely employed in supervising and controlling critical infrastructures, with manufacturing systems that incorporate industrial robotic arms being a prominent example. The increasing adoption of ubiquitous computing technologies in these systems has led to benefits such as real-time monitoring, reduced maintenance costs, and high interconnectivity. This adoption has also brought cybersecurity vulnerabilities exploited by adversaries disrupting manufacturing processes via manipulating actuator behaviors. Previous incidents in the industrial cyber domain prove that adversaries launch sophisticated attacks rendering network-based anomaly detection mechanisms insufficient as the “physics” involved in the process is overlooked. To address this issue, we propose an IoT-based cyber-physical anomaly detection system that can detect motion-based behavioral changes in an industrial robotic arm. We apply both statistical and state-of-the-art machine learning methods to real-time Inertial Measurement Unit data collected from an edge development board attached to an arm doing a pick-and-place operation. To generate anomalies, we modify the joint velocity of the arm. Our goal is to create an air-gapped secondary protection layer to detect “physical” anomalies without depending on the integrity of network data, thus augmenting overall anomaly detection capability. Our empirical results show that the proposed system, which utilizes 1D convolutional neural networks, can successfully detect motion-based anomalies on a real-world industrial robotic arm. The significance of our work lies in its contribution to developing a comprehensive solution for ICPS security, which goes beyond conventional network-based methods.