28579 research outputs found
Sort by
Adaptive Federated Learning Defense Against Byzantine Attacks and Concept Drift in IIoT
Get PDFFederated Learning (FL) enables collaborative intrusion detection in Industrial Internet of Things (IIoT) environments without compromising data privacy. However, FL systems face critical challenges from Byzantine attacks, where malicious clients send poisoned model updates, and concept drift, where data distributions evolve over time. Existing defenses typically force a trade-off between security and efficiency, employing either computationally expensive robust aggregation methods or fast but vulnerable approaches. This paper proposes an adaptive defense framework that dynamically responds to the threat landscape using lightweight statistical detection mechanisms. The system defaults to efficient Federated Averaging (FedAvg) aggregation and switches to robust methods only when attacks or drift are detected. We validated the framework through experiments on the Edge-IIoTset dataset and real-world deployment on five Raspberry Pi 3B devices. Results show the adaptive approach maintains F1=0.828 under 40% malicious clients and remains robust up to the 50% threshold
FocusViT: Faithful Explanations for Vision Transformers via Gradient-Guided Layer-Skipping
Get PDFVision Transformers (ViTs) have emerged as powerful alternatives to CNNs for various vision tasks, yet their token-based, attention-driven architecture makes interpreting their predictions challenging. Existing explainability methods, such as Grad-CAM and Attention Rollout, either fail to capture hierarchical semantic information or assume atten tion directly reflects importance, often lead- ing to misleading explanations. We propose FocusViT, a novel explainability framework that integrates gradient-weighted attention attribution with dynamic, faithfulness-driven layer aggregation. By fusing attention maps with class-specific gradients and introducing per-head dynamic weighting, FocusViT highlights not only where the model attends but also how sensitive the prediction is to those attentions. Furthermore, our adaptive layer-skipping strategy ensures that only semantically meaningful layers contribute to the final explanation, enhancing both faithfulness and clarity. Extensive quantitative and qualitative evaluations on diverse benchmarks demonstrate that FocusViT improves over existing methods in faithfulness and sparsity, achieving competitive robustness and class sensitivity, and provides sharper, more reliable visual explanations for ViTs. The official implementation is publicly available at: https://github.com/game-sys/focusvit- aistats2026.gi
Peace treaties’ environmental provisions and conflict recurrence
Get PDFA fundamental idea in the environmental peacebuilding literature is that environmental cooperation can serve as a catalyst for broader collaboration and, in turn, reduce the risk of renewed conflict. While this claim has long been asserted, only a few empirical studies have examined it in cross-country and comparative analyses. This article adds to these works by providing a comprehensive quantitative analysis of conflict recurrence data and provisions for environmental cooperation in over 1,000 peace treaties since the end of the Cold War. The central finding is that environmental clauses in such agreements are associated with a significantly lower risk of conflict recurrence. This result not only provides systematic empirical evidence supporting the environmental peacebuilding thesis but also demonstrates that environmental cooperation is not merely a symbolic gesture. Rather, it plays a vital, substantive role in establishing sustainable post-conflict stability and durable resolution
Long Covid: a longitudinal perspective of cognitive impairment and symptom trajectories in a working age sample
Get PDFSARS-CoV-2, has been shown to induce inflammation in the hippocampus which may lead to prolonged memory deficits in people with Long Covid (pwLC). While Long Covid (LC) is estimated to affect over 65 million people worldwide, its long-term cognitive and functional sequalae remain poorly understood. Using a mixed-methods longitudinal design, this thesis investigated the symptom trajectories of LC. It also examined the lived experiences of pwLC, integrating qualitative insights alongside quantitative data. People of working age (N = 68), most of whom (81%) had a formal LC diagnosis, completed sub-tests of the BIRT Memory and Information Processing Battery (BMIPB-II), a word categorisation and recognition task (Addante et al., 2012) and health surveys at three time points, from March 2023 to March 2025 with an average of 19 months between the first and third assessment (SD = 1.41). Qualitative data from open-ended survey responses and semi-structured interviews, were examined using thematic analysis to capture lived experiences. Finally, a study protocol was designed to explore cortical dynamics of memory encoding and retrieval, to further delineate the neurophysiological underpinnings of the condition. Persistent impairments in memory were observed for up to 60 months since LC diagnosis. Verbal learning and retention and speed of information processing were impaired on all assessments. Visual memory, semantic categorisation accuracy and speed were preserved. Participants self-reported up to 34 LC symptoms. The most frequent were fatigue, memory and concentration problems. PwLC demonstrated elevated fatigue compared to controls, which correlated with low quality of life. Linear mixed modelling showed changes over time for some of the cognitive performance measures, but not on all other measures. Lived experience perspectives demonstrated the complex and multifaceted challenges of living with LC. The findings from this thesis evidence the persistent cognitive and functional deficits in pwLC, and may inform the development of better diagnosis and support pathways for pwLC
A psychoanalytic exploration of the emotional experience of breastfeeding mothers who have experienced difficulties feeding their baby with tongue-tie. An interpretative phenomenological analytic study
Get PDFBackground: Infant tongue-tie or ankyloglossia is a relatively minor and easily treated condition associated with breastfeeding complications which have been linked with consequential emotional maternal distress. Research Aim: To gain insight into lived emotional experiences of mothers who have experienced difficulties breastfeeding their tongue-tied baby. Method: This was a qualitative study of four participants who had breastfed a baby with tongue-tie (aged 3-12 months) using a single Free Association Narrative Interview (FANI) - aimed at making available inner-world and unconscious experiences. Data analysis by Interpretative Phenomenological Analysis (IPA) focussed on unique idiographic aspects of individual lived experience leading to identified recurrent group themes. Within the IPA framework, aspects of FANI and Reverie Research methods were included, adopting psychoanalytic concepts, tools and understanding to promote and deepen the unconscious experience made available for understanding within the hermeneutic IPA process. Results: Four main Group Experiential Themes were identified: “The psychological and emotional complexity as a breastfeeding mother”; “The hope of connection and the pain of disconnection”; “Internal relationships with dependency and care in interaction with lived external experiences”; and “Alone and stuck in an unbearable situation – reaching breaking point”. Conclusions: Results suggest breastfeeding is both physical and emotional, with success linked with feelings of maternal confidence and capacity, and difficulties with the reverse. Maternal emotional experiences of infant tongue-tie were inextricably linked with pregnancy and delivery, and past and current lived experiences. Professional care was profoundly experienced - positively and negatively. If left feeling unsupported, participants reached a state of desperation disrupting their sense of connection with their baby and with the support around them. By contrast, timely, specialist attentive care and emotional containment promoted emotional recovery and a reparative breastfeeding relationship with their baby
Post-Quantum Protected Federated Learning with Explainable and Adaptive Intelligence for Smart City Transportation
No full textExisting AI-powered Intelligent Transportation Systems (ITS) have limitations in scalability, privacy, and vulnerability to cyberattacks, as well as a lack of transparency in decision-making. In this work, we present a hybrid framework based on Post-Quantum-protected Federated Learning, a lightweight CNN-Transformer model, LIME explanations, and a local model, achieving a loss of 0.02% and a validation accuracy of 98%. At the boundary, congestion is determined using CityFlowV2 traffic camera feeds, which are based on Federated Learning, a distributed training framework that does not require sharing raw data, and the architecture is privacy-respectful. Reinforcement learning trained on OpenStreetMap road networks in Los Angeles coordinates rerouting plans in a simulated environment at the global level, and SHAP provides an explanation of the decision. The Federated aggregation retained accuracy at the zone level, exceeding 97%. Furthermore, this affirms its strength. CRYSTALS-Kyber is used to encrypt V2I and V2V communications, ensuring they are resistant to attacks in the quantum era
Commercial fishing amplifies impacts of increasing temperature on predator-prey interactions in marine ecosystems
Get PDFPredator-prey interactions determine food web structure, energy flux, and ecosystem stability. Increasing temperatures and commercial fishing both alter body size distributions that underpin predator-prey interactions, but empirical evidence of their individual and combined effects is limited. We study how the predator to prey body mass ratio (PPMR) changes as a function of temperature and fishing effort in over 50,000 predator stomachs collected across the Northeast Atlantic over 35 years. PPMR increases with temperature, an effect that is exacerbated by greater fishing effort, driven by intraspecific decreases in prey body mass in heavily fished areas. To compensate for smaller prey (both within and across species) in warmer waters and areas of high fishing, predators target the largest prey available to them, but this is insufficient to alter the community-wide increase in PPMR. Higher PPMR is associated with weaker trophic interactions that dampen strong oscillatory dynamics but could also reduce energy transfer efficiency within ecosystems, both of which can affect ecosystem stability. These results could help underpin ecosystem-based management and sustainable fisheries by providing estimates of how future climate warming might interact with fishing to affect energy flux through marine food webs
Strongly Typed Cartesian Genetic Programming and its applications
Get PDFGenetic Programming (GP) and its graph-based variant, Cartesian Genetic Programming (CGP), have proved highly effective machine-learning frameworks that can evolve human-readable programs that can match or exceed hand-crafted solutions across many tasks. Strongly Typed Cartesian Genetic Programming (ST–CGP) extends CGP by assigning explicit data-types to every input, output and operator, and allowing features to have varying arities. These type constraints prune infeasible regions of the search space while preserving CGP’s directed-acyclic-graph representation and single-row genome, leading to faster, semantically correct evolution. Unlike standard CGP, ST–CGP also introduces two forms of crossover—full two-point recombination and “genetic rewiring”—providing a second source of variation that is rarely available in conventional CGP frameworks. Because operators are typed, ST–CGP can be rapidly retargeted: numeric, boolean and higher-level domain primitives (e.g. OpenCV filters) can coexist in a single run, enabling one framework to span diverse problem domains. This versatility is illustrated in this thesis by three application areas. In computer vision, ST–CGP evolved segmentation, detection and classification pipelines that solved benchmark object-sorting problems and achieved convolutional-neural-network-level accuracy on a 27,000-image malaria-cell dataset with far smaller training sets and CPU-only resources. In agriculture, it classified field parcels into low, high and reference yield zones using laboratory soil measurements with competitive accuracy and markedly low variance relative to traditional models. Finally, it learned predictive models mapping five-minute VOC gas “fingerprints” from an electronic-nose sensor to multiple soil health indicators, delivering laboratory-grade predictions, an application which has now been adopted by UK agronomists in commercial practice. Collectively, these results demonstrate that the combination of strong typing, an enriched operator palette and novel crossover elevates CGP to a general-purpose, interpretable evolutionary programming system capable of tackling data-rich tasks from medical imaging to environmental sensing within a single unified framework
CEOs’ Early-Life Disaster Experiences and Corporate Hedging Activities
Get PDFWe study how traumatic experiences in childhood influence CEOs’ risk preferences and corporate financial hedging decisions. Based on a sample of U.S. public firms from 1993 to 2020, we document a positive relation between CEOs’ early-life disaster experiences and the likelihood of firms using financial derivatives. We also find that the interactive impact of disaster experiences and financial hedging on firm value is negative, suggesting that early-life disaster experiences increase the gap between CEOs’ and shareholders’ risk preferences, potentially leading to conflicts of interest. Furthermore, our cross-sectional analysis shows that the positive relation between disaster experiences and financial hedging is more pronounced in firms with weaker corporate governance, fewer financial constraints, and higher firm-specific risk. Our findings suggest that corporate boards and regulators should maintain active oversight of corporate risk management practices, especially when early-life disaster experiences are known to influence a CEO’s risk preferences
Communication modality, authenticity, and continuance usage intention of GenAI chatbots: A media richness theory perspective
No full textGenerative AI chatbots are increasingly popular for tourist destination information searches. However, how communication modalities (text vs. voice), interaction styles (social vs. task-oriented), destination types (hedonic vs. utilitarian), and their interactions contribute to users’ perceptions and continuance usage intention remains unclear. Building on media richness theory, this study used a sequential explanatory mixed-methods and multi-study research design, with four scenario-based experiments to examine how the above factors affect tourists’ perceived authenticity of GenAI chatbots and continuance usage intention, and a focus group study to validate and contextualize findings from experimental studies. The results indicated that voice communication evokes a higher level of GenAI chatbots’ authenticity. Individuals’ perceived authenticity of GenAI chatbots is positively associated with their continuance usage intention. Destination type was a significant moderator, with voice modality enhancing authenticity more for hedonic than utilitarian destinations. The moderating roles of destination type and interaction style are clarified, shedding new light on destination marketing theories and practices