José Cecílio

Mixture-of-Experts (MoE) architectures have recently been investigated for safety-critical perception tasks due to their computational efficiency and scalability. Condition-supervised MoE models support context-aware autonomous driving perception by using a router (gate network) to assign adaptive weights to condition-specialized expert branches. However, this routing mechanism also introduces a distinct attack surface, since adversarial perturbations that alter the routing distribution can change expert utilization and degrade the fused output even when the expert branches remain object-discriminative under clean inputs. Existing adversarial robustness studies on MoE models remain limited for autonomous driving perception, as they rarely report robustness degradation at the router, expert, and fused output levels. They also provide limited or no evidence of vulnerability under black-box attacks targeting the routing module rather than the complete MoE model. This paper presents a routing-aware adversarial evaluation framework that integrates several black-box attack methods under a common l∞ perturbation budget and introduces a vulnerability diagnostic protocol that decomposes adversarial degradation across the routing stage, the individual expert output heads, and the end-to-end fused prediction head. Empirical results on the BDD100K dataset show that routing disruption accounts for the largest observed degradation of robustness under the evaluated attacks. Supplementary evaluation on the nuImages dataset shows that the same vulnerability pattern persists when the auxiliary taxonomy changes from weather, scene, and time-of-day to location, illumination, and ego-motion.

The use of drones in naval applications, both military and civilian, has raised concerns about their security. Their protection should not be limited to the physical domain of kinetic warfare or the electromagnetic domain of electronic warfare. It must also take into account cyber threats, given the extensive reliance of drones on computational and communication systems. To support the study of cyber threats and the development of security solutions for drone systems, this work presents a simulation platform (CyDronE) that brings together not only elements of the cyber domain but also properties of the physical domain. CyDronE is implemented on a containerized Robot Operating System 2 (ROS 2) framework, which executes the drones’ onboard computational processes. Gazebo is integrated as the physics-based simulation engine, providing high-fidelity modeling of the drones’ flight dynamics and environmental interactions during the experiments. This approach enables a consistent assessment of how cyber actions influence the physical behavior of drones. System components are defined through declarative YAML configurations and instantiated into deterministic, per-swarm deployments, allowing multiple isolated swarm instances to execute concurrently on the same host without mutual interference. To evaluate the system, we conduct experiments involving a swarm of drones operating under adversarial conditions. Specifically, we implement a controlled UDP flood attack using dedicated attacker containers that generate traffic directed at the same IP addresses and ports employed by DDS/RTPS communication. The simulations demonstrate that flooding can hinder the timely delivery of legitimate state updates, resulting in degraded vehicle motion, transitioning from continuous trajectories to prolonged hovering with intermittent advancement. These experimental results demonstrate the system’s capability to support cyber–physical assessments of cybersecurity activities in drone applications. Additionally, the assessment of the computational resources consumed by CyDronE suggests its good scalability.

Autonomous driving perception must maintain reliable object presence prediction under heterogeneous operating conditions, where visual evidence for the same object category can vary substantially with weather, scene structure, illumination, location, and motion state. Conventional shared-backbone perception models learn a single representation for all inputs, potentially underutilizing condition-specific visual cues. Standard mixture-of-experts models offer conditional computation, but their routing behavior is often learned from task-prediction gradients, which can lead to unstable expert allocation, weak specialization, or routing collapse. This paper proposes Condition-Supervised Adaptive Gating Mixture-of-Experts (CSA-GMoE), a condition-aware MoE framework for image-level multi-label object-presence prediction in autonomous driving scenes. CSA-GMoE uses dataset-provided driving-context annotations to supervise expert specialization and trains a gradient-decoupled managing network through structured input-to-expert alignment. Expert-level object-presence heads preserve object-discriminative capacity, while masked auxiliary supervision enables handling incomplete contextual annotations without discarding primary-task samples. Experiments on BDD100K show that CSA-GMoE achieves stable non-degenerate routing, preserves expert-level object discrimination, and obtains strong fused prediction performance, with a macro F1-score of 0.838 , mAP of 0.920, ECE of 0.109 , and inference time of 1.788-ms per image. Additional experiments on NuImages show that the framework remains applicable when the auxiliary condition taxonomy changes from weather-scene-time to location-illumination-motion. These results indicate that contextual driving annotations can provide structured supervision for interpretable and condition-aware expert specialization in autonomous-driving perception.

Household appliances, particularly washing machines and tumble dryers, account for a significant share of residential energy consumption. While energy labelling aims to encourage the adoption of more efficient appliances, actual energy and water usage are highly influenced by user behaviour and operating conditions. The user behaviour is often overlooked in existing datasets. To address this gap, the household Laundry Appliance Resource Consumption and Operation dataset (LARCO) is presented: a comprehensive, open-access dataset capturing detailed operational and environmental data from washing machines and dryers under laboratory and household conditions. The dataset includes multivariate time-series recordings at 1 Hz (and higher for vibrations and audio), covering a range of load sizes, programme settings, and ambient conditions. Laboratory data provide ground-truth measurements under controlled conditions, while household data capture real-world usage variability. Each sample includes energy, water, temperature, and humidity measurements, associated with the corresponding metadata. LARCO enables studies on appliance efficiency, user behaviour modelling, and personalised energy feedback. It supports the development of more effective strategies to bridge the gap between labelled and actual appliance performance, promoting more sustainable energy use.

As the number of IoT devices grows, ensuring the secure validation and processing of the data they generate has become critical. This challenge becomes even more pronounced for mobile nodes, such as vehicles, which must continually reconnect to new edge servers while in motion. This work proposes a decentralized architecture for connecting IoT devices to edge servers, enabling secure data delivery to applications while minimizing overhead and ensuring trustworthy handovers between edge servers. A fundamental concern is the integrity of edge servers, as they may be compromised or exhibit malicious behavior. To address this, the proposed architecture relies on an external verifier service to continuously verify the integrity of edge servers.To demonstrate its feasibility, two prototypes were implemented using distinct consensus technologies and evaluated under realistic conditions. The first prototype, based on BFT-SMaRt, achieved lower latency and higher throughput but required dedicated, proprietary infrastructure and lacked global auditability. In contrast, the second prototype, leveraging an existing blockchain network, provides complete auditability and decentralization without proprietary infrastructure, though at the cost of higher latency. Experimental results confirm that both approaches deliver strong security guarantees, with trade-offs between performance and transparency, validating the architecture’s suitability for dynamic IoT environments.