The limited processing capacity of computing equipment that is usually mounted on Unmanned Aerial Vehicles (UAVs) often prevents real-time execution of computer vision tasks, such as image segmentation. In this paper, we introduce SqueezeSlimU-Net (SSU-Net), an adaptive and efficient deep learning (DL) model designed to enhance UAV capabilities in performing complex image segmentation tasks under resource constraints, thereby advancing real-time UAV vision – a crucial technology in fields such as precision agriculture. SSU-Net combines benefits of three specialized DL architectures: the semantic segmentation capabilities of the U-Net architecture, the computational efficiency of SqueezeNet's fire modules, and the dynamic adaptability of Slimmable Neural Networks (SNNs). This integration allows SSU-Net to adjust its network width in real-time, thus striking the balance between inference accuracy and computational load based on the operational parameters such as task requirements and UAV's battery life. To validate SSU-Net's efficacy, we applied it to a weed detection task using two UAV-collected datasets and tested it on an edge computing platform for UAVs. Our experiments show that SSU-Net can reduce inference energy consumption by up to 65% with only a minimal 2% reduction in accuracy. A comparative evaluation with other state-of-the-art DL image segmentation approaches shows that SSU-Net achieves on par weed detection performance while requiring significantly fewer model parameters. Additionally, SSU-Net outperforms state-of-the-art network pruning techniques in balancing accuracy and resource usage. Timing benchmarks show SSU-Net fostering real-time weed detection even on low-resource UAVs, making it ideal for UAV remote sensing applications.

The shortcomings of the traditional password-based authentication mechanism are becoming increasingly apparent as we transition from "one user - one device" to a richer "multiple users - multiple devices" computing paradigm. The currently dominant research direction focuses on on-device biometrics, which require sensitive information, such as images of a user's face, to be constantly streamed from a single recording source, often the device on which a user is getting authenticated. Instead, in this work we explore the possibilities offered by heterogeneous devices that opportunistically collect non-sensitive data in smart environments. We construct an IoT testbed in which we gather data pertaining to a person's movement in space, interaction with certain physical objects, PC terminal usage, and keyboard typing, and construct machine learning models capturing the person's behaviour traits. We commence our examination with models constructed from data sensed during a previously-completed task run and with such models we achieve up to 68% user identification accuracy (c.f. 7% baseline) among up to 20 individuals. Taking into account the limits of behaviour persistence we then revise our approach to continuously refine the model with the most recently sampled sensor data. This method allows us to achieve 99.3% user verification accuracy and successfully prevent a session takeover attack within 12 seconds with less than 1% of false attack detection.

Cyber Threat Intelligence (CTI) reports provide information about emerging and current cyber threats, and their analysis is key for adopting appropriate countermeasures. Reports are typically in the form of long texts from which cybersecurity analysts extract essential elements and translate them into actionable steps. To summarise and share the findings of this analysis, sentences in the reports are often labelled with MITRE ATT&CK techniques that yield a better description of the identified attack patterns. However, this task can be very time-consuming and prone to both errors and biases of analysts. In the literature, there have been some attempts to automate this process. Most commonly, researchers apply different pre-processing steps on the initial reports and then apply classification techniques, including approaches based on large language models (LLMs). Considering that reports are written in natural language, in this paper, we present an approach that relies entirely on LLMs and seeks to minimise preprocessing of reports and other human intervention, if not to replace, at least to ease the task of the analysts. We evaluate our approach on a real-world CTI report and an extensive dataset of MITRE-labelled sentences and reduce the number of potentially suitable techniques by up to 33 $$\times $$ × while retaining ground truth labels in up to 94.29% of the sentences.

The 2022--2023 food crises and the ongoing human population growth make the efficient use of the available agricultural land a pressing matter. However, weeds present a major obstacle towards efficient land use, and cause up to 40% yield loss in all major crops, leading to more than $100 billion USD annual global economic loss. Camera equipped unmanned aerial vehicles (UAVs) represent an attractive tool for automatic weed detection. Yet, high computation and energy requirements of deep learning models restrict automatic real-time inference to expensive high-end UAVs, preventing wider adoption of this promising solution for weed detection. In this work we present AgriAdapt, a solution for lightweight on-UAV weed detection that is based on novel slimmable U-Net neural architecture for weed detection. A defining property of AgriAdapt is its adaptability to operating conditions - the resource usage can be scaled dynamically according to the needs for battery preservation, input difficulty, or other factors. In this paper we present preliminary experiments on a newly-collected weeds dataset and detailed assessment of resource savings enabled by AgriAdapt.

The global food crises and growing population necessitate efficient agricultural land use. Weeds cause up to 40% yield loss in major crops, resulting in over USD 100 billion in annual economic losses. Camera-equipped UAVs offer a solution for automatic weed detection, but the high computational and energy demands of deep learning models limit their use to expensive, high-end UAVs. In this paper, we present a low-cost UAV system built from off-the-shelf components, featuring a custom-designed on-board computing system based on the NVIDIA Jetson Nano. This system efficiently manages real-time image acquisition and inference using the energy-efficient Squeeze U-Net neural network for weed detection. Our approach ensures the pipeline operates in real time without affecting the drone’s flight autonomy. We also introduce the AgriAdapt dataset, a novel collection of 643 high-resolution aerial images of salad crops with weeds, which fills a key gap by providing realistic UAV data for benchmarking segmentation models under field conditions. Several deep learning models are trained and validated on the newly introduced AgriAdapt dataset, demonstrating its suitability for effective weed segmentation in UAV imagery. Quantitative results show that the dataset supports a range of architectures, from larger models such as DeepLabV3 to smaller, lightweight networks like Squeeze U-Net (with only 2.5 M parameters), achieving high accuracy (around 90%) across the board. These contributions distinguish our work from earlier UAV-based weed detection systems by combining a novel dataset with a comprehensive evaluation of accuracy, latency, and energy efficiency, thus directly targeting deep learning applications for real-time UAV deployment. Our results demonstrate the feasibility of deploying a low-cost, energy-efficient UAV system for real-time weed detection, making advanced agricultural technology more accessible and practical for widespread use.

Identifying humans based on their behavioural patterns represents an attractive basis for access control as such patterns appear naturally, do not require a focused effort from the user side, and do not impose the additional burden of memorising passwords. One means of capturing behavioural patterns is through passive sensors laid out in a target environment. Thanks to the proliferation of the Internet of Things (IoT), sensing devices are already embedded in our everyday surroundings and represent a rich source of multimodal data. Nevertheless, collecting such data for authentication research purposes is challenging, as it entails management and synchronisation of a range of sensing devices, design of diverse tasks that would evoke different behaviour patterns, storage and pre-processing of data arriving from multiple sources, and the execution of long-lasting user activities. Consequently, to the best of our knowledge, no publicly available datasets suitable for behaviour-based authentication research exist. In this brief article, we describe the first multimodal dataset for behavioural authentication research collected in a sensor-enabled IoT setting. The dataset comprises of high-frequency accelerometer, gyroscope, and force sensor data collected from an office-like environment. In addition, the dataset contains 3D point clouds collected with wireless radar and electroencephalogram (EEG) readings from a wireless EEG cap worn by the study participants. Within the environment, 54 volunteers conducted 6 different tasks that were constructed to elicit different behaviours and different cognitive load levels, resulting in a total of 16 h of multimodal data. The richness of the dataset comprising 5 different sensing modalities, a variability of tasks including keyboard typing, hand gesturing, walking, and other activities, opens a range of opportunities for research in behaviour-based authentication, but also the understanding of the role of different tasks and cognitive load levels on human behaviour.