AI at the network edge promises real-time perception and decision-making in safety-critical domains such as aerial robotics, autonomous vehicles, and 5G-enabled infrastructures. Yet, operating under resource constraints, dynamic, and adversarial conditions exposes edge AI systems to fragility, inefficiency, and security risks that threaten their safe operation. GuardAI, a Horizon Europe project, introduces a framework for resilient and trustworthy edge AI that unites three pillars: adversarial robustness, context-enhanced inference, and security-by-design. Initial project results include a diffusion-based adversarial purification framework optimized for real-time operation, lightweight deep unrolling architectures for LiDAR super-resolution with built-in outlier removal, and robust uncertainty quantification modules to improve confidence calibration. It further develops a context-enhanced inference engine that integrates visual, spatial, and operational context across multi-agent systems, and a risk-aware defense recommender that autonomously selects mitigation strategies based on evolving threat landscapes. Through representative Use Cases, covering monitoring with Unmanned Aerial Vehicle, decentralized 5G network analytics, and secure perception in connected autonomous vehicles, GuardAI demonstrates how robust and adaptive AI can be achieved within stringent edge constraints. Together, these technologies lay the groundwork for a new generation of secure, context-aware, and certifiable AI systems that can be trusted to operate autonomously in the physical world.

Long-standing data sparsity and cold-start constitute thorny and perplexing problems for the recommendation systems. Cross-domain recommendation as a domain adaptation framework has been utilized to effectively address these challenging issues, by exploiting information from multiple domains. In this study, an item-level relevance cross-domain recommendation task is explored, where two related domains, that is, the source and the target domain contain common items. Additionally, a user-level relevance scenario is considered, where the two related domains contain common users. In light of these scenarios, two novel coupled autoencoder-based deep learning methods are proposed for cross-domain recommendation. The first method aims at simultaneously learning a pair of autoencoders in order to reveal the intrinsic representations in the source and target domains, along with a coupled mapping function to model the non-linear relationships between these representations. The second method is derived based on a new joint regularized optimization problem, which employs two autoencoders to generate in a deep and non-linear manner the user and item-latent factors, while at the same time a data-driven function is learned to map the latent factors across domains. Extensive numerical experiments are conducted illustrating the superior performance of our proposed methods compared to several state-of-the-art cross-domain recommendation frameworks.

This work tackles the challenge of enhancing low-resolution LiDAR sensors for SLAM applications through a novel Deep Unrolling-based Super-Resolution (SR) model. We integrate an outlier removal module to ensure structural integrity while maintaining real-time performance. By leveraging a model-based optimization approach, our method efficiently reconstructs high-resolution point clouds while minimizing computational overhead. The proposed SR model is evaluated within a LiDAR SLAM framework, demonstrating significant improvements in pose estimation accuracy and efficiency compared to state-of-the-art SR methods.

The critical perception capabilities of EdgeAI systems, such as autonomous vehicles, are required to be resilient against adversarial threats, by enabling accurate identification and localization of multiple objects in the scene over time, mitigating their impact. Single-agent tracking offers resilience to adversarial attacks but lacks situational awareness, underscoring the need for multi-agent cooperation to enhance context understanding and robustness. This paper proposes a novel mitigation framework on 3D LiDAR scene against adversarial noise by tracking objects based on least-squares graph on multi-agent adversarial bounding boxes. Specifically, we employ the least-squares graph tool to reduce the induced positional error of each detection’s centroid utilizing overlapped bounding boxes on a fully connected graph via differential coordinates and anchor points. Hence, the multi-vehicle detections are fused and refined mitigating the adversarial impact, and associated with existing tracks in two stages performing tracking to further suppress the adversarial threat. An extensive evaluation study on the real-world V2V4Real dataset demonstrates that the proposed method significantly outperforms both state-of-the-art single and multi-agent tracking frameworks by up to 23.3% under challenging adversarial conditions, operating as a resilient approach without relying on additional defense mechanisms or training parameters.

In this paper we propose a methodology combining Federated Learning (FL) with Cross-view Image Geo-localization (CVGL) techniques. We address the challenges of data privacy and heterogeneity in autonomous vehicle environments by proposing a personalized Federated Learning scenario that allows selective sharing of model parameters. Our method implements a coarse-to-fine approach, where clients share only the coarse feature extractors while keeping fine-grained features specific to local environments. We evaluate our approach against traditional centralized and single-client training schemes using the KITTI dataset combined with satellite imagery. Results demonstrate that our federated CVGL method achieves performance close to centralized training while maintaining data privacy. The proposed partial model sharing strategy shows comparable or slightly better performance than classical FL, offering significant reduced communication overhead without sacrificing accuracy. Our work contributes to more robust and privacy-preserving localization systems for autonomous vehicles operating in diverse environments.

This paper proposes a novel localization framework based on collaborative training or federated learning paradigm, for highly accurate localization of autonomous vehicles. More specifically, we build on the standard approach of KalmanNet, a recurrent neural network aiming to estimate the underlying system uncertainty of traditional Extended Kalman Filtering, and reformulate it by the adapt-then-combine concept to FedKalmanNet. The latter is trained in a distributed manner by a group of vehicles (or clients), with local training datasets consisting of vehicular location and velocity measurements, through a global server aggregation operation. The FedKalmanNet is then used by each vehicle to localize itself, by estimating the associated system uncertainty matrices (i.e, Kalman gain). Our aim is to actually demonstrate the benefits of collaborative training for state estimation in autonomous driving, over collaborative decision-making which requires rich V2X communication resources for measurement exchange and sensor fusion under real-time constraints. An extensive experimental and evaluation study conducted in CARLA autonomous driving simulator highlights the superior performance of FedKalmanNet over state-of-the-art collaborative decision-making approaches, in localizing vehicles without the need of real-time V2X communication.

In this paper, a novel technique for the hyperspectral image deconvolution problem is developed. First, considering the highly ill-posed nature of the examined problem, it is imperative to incorporate proper priors (regularizers) to capture the strong spectral and spatial dependencies of the hyperspectral images. Then, in light of this, a novel optimization problem is proposed by employing a convolutional neural network to act as a regularizer, which is learnt to reflect the properties of the signals of interest. To solve the proposed optimization problem, we use the half quadratic splitting methodology, thus designing an efficient iterative solver (iteration map). Based on the Deep Equilibrium (DEQ) modeling, which aims to express the proposed iterative solver as an equilibrium (fixed-point) computation, a highly interpretable deep learning-based network is derived, which can be trained endto-end. Extensive numerical results using two publicly available datasets illustrate that the proposed method markedly outperforms other state-of-the-art approaches.

Considering the high cost of high-resolution Lidar sensors, in this work, a novel Lidar super-resolution method is proposed to improve the performance on numerous autonomous vehicle perception tasks, including that of a Lidar odometer. Specifically, we propose a regularized optimization problem employing a learnable regularizer (neural network) to capture the properties of the data. To efficiently solve this problem, a deep unrolling methodology is proposed, thus forming an interpretable and well-justified deep architecture. Extensive experiments on a real-world lidar odometry application highlight that the proposed model exhibits both superior performance as well as a significantly reduced number of trainable parameters i.e., 99.75% less parameters, as compared to other deep learning methods. The source code used for this work can be found at our repository: repository.

This demo presents a novel Deep Federated Unrolling (FL-DU) super-resolution (SR) approach for enabling low-resolution Lidar-based Simultaneous Localization and Mapping (SLAM) solutions. The proposed system enhances the accuracy of low-cost Lidar sensors via novel explainable by design neural networks and by enabling collaboration between individual vehicles during learning, thereby minimizing the need for costly high-resolution Lidar sensors, leading to significant cost reductions without affecting the SLAM accuracy. Our demo is available on https://www.youtube.com/watcn?v=Fp_nBrD6NiY.

In this study, the problem of computing a sparse representation of multi-dimensional visual data is considered. In general, such data e.g., hyperspectral images, color images or video data consists of signals that exhibit strong local dependencies. A new computationally efficient sparse coding optimization problem is derived by employing regularization terms that are adapted to the properties of the signals of interest. Exploiting the merits of the learnable regularization techniques, a neural network is employed to act as structure prior and reveal the underlying signal dependencies. To solve the optimization problem Deep unrolling and Deep equilibrium based algorithms are developed, forming highly interpretable and concise deep-learning-based architectures, that process the input dataset in a block-by-block fashion. Extensive simulation results, in the context of hyperspectral image denoising, are provided, which demonstrate that the proposed algorithms outperform significantly other sparse coding approaches and exhibit superior performance against recent state-of-the-art deep-learning-based denoising models. In a wider perspective, our work provides a unique bridge between a classic approach, that is the sparse representation theory, and modern representation tools that are based on deep learning modeling.

In multidimensional times series generated by sensor recordings of multiple dispersed IoT edge devices, missing measurements are commonplace due to sensing or communication failures, considered a thorny and perplexing problem in a wide range of Industry 4.0 applications. Existing studies for time series imputation focus on developing centralized deep learning approaches, which require massive amounts of data to be uploaded to a central server with adequate computational and power resources for the training of the models, since these approaches are unsuitable for edge and IoT devices characterized by limited computation resources. Different from the current literature, in this study, the time series imputation problem is studied from a federated learning perspective, which is able to surmount the above difficulties. In particular, a novel federated learning approach is proposed, assuming different IoT devices with varying sensing and computational capabilities, that trade-off accuracy with computational/communication/sensing complexity and minimize the operations that need to be performed during training and inferences phase. Furthermore, considering that the main computations are performed on the edge, where the IoT edge devices have limited computational capabilities and power resources, a lightweight yet effective autoencoder-based model is employed to address the examined problem, modified properly to capture the temporal dependencies of the time series data. Extensive evaluation studies with two open datasets have shown that both approaches minimize the data exchanges the need to be made for outperforming centralized approaches in the presence of limited training data.

Given two datasets that belong to different feature spaces and both correspond to the same underlying phenomenon, the scope of coupled dictionary learning is to compute two dictionaries, one for each dataset, so that each dataset is approximated using the respective dictionary but the same sparse coding matrix. In this work, the focus is on a particular, yet widespread, form of this problem in which the datasets correspond to slowly varying (piece-wise smooth) signals, and the measurements contain severe noise. A novel coupled dictionary learning technique is developed by including a suitable total-variation-based regularization term in the cost function. Furthermore, exploiting the smoothness of the datasets, new fast sparse coding algorithms are derived. The new techniques achieve effective modeling of the smooth signal and significantly alleviate the effects of noise. Finally, extensive simulation results for the problem of spectral super-resolution of hyperspectral images are provided, demonstrating the performance improvements offered by the derived techniques.