We present BuildTree, a deterministic silent self-stabilizing distributed algorithm that constructs a clustered spanning tree under the unfair daemon. From any arbitrary configuration, the system converges to a legal state in which nodes are organized into bounded-diameter clusters connected by a spanning tree; once stabilized, no further actions occur. The clustered structure implies a provable bound on the average stretch of the resulting tree as a function of the cluster diameter, yielding polynomially bounded average stretch for standard parameters. The stabilization time is O(n+Δ·η) rounds, where n is the number of nodes, Δ the maximum cluster size, and η the number of clusters. To our knowledge, this is the first silent self-stabilizing construction of clustered spanning trees with bounded stretch guarantees.
Vector processors efficiently run High-Performance Computing (HPC) workloads. Modern commercial vector architectures feature large physical vector register files that have hundreds of kilobytes of storage capacity. This large storage capacity brings many data reuse opportunities at the register file level that can potentially deliver large performance benefits. This paper presents Registers-as-a-Cache (RaaC), a light-weight extension of the vector register file that makes it possible to serve load demand requests using data present in vector registers, which dramatically reduces memory access latency. RaaC enables data reuse at the register level, effectively avoiding a large portion of memory references to the memory hierarchy. When deployed on top of a state-of-the-art out-of-order vector processor, RaaC obtains 1.26 × speed-up, serves 55.3% of the memory references by reusing data already stored in registers, reduces the average memory access latency by 74.6%, and improves energy-to-solution by 14.9%. RaaC outperforms previous microarchitecture techniques that exploit data reuse at the register file by 1.24 × .
Federated Domain Generalization (FedDG) is a challenging problem that aims to learn a global model that can generalize to the unseen target domain while preserving privacy. Previous state-of-the-art (SOTA) DG methods have struggled to achieve this goal due to the requirement of centralized training. Existing FedDG methods either share domain-specific information across clients, risking privacy leakage, or do not explicitly disentangle domain-invariant content from domain-specific style, limiting generalization to unseen domains. These limitations call for a framework that achieves feature-level disentanglement under FL privacy constraints without sharing client data. In this paper, we reconsider the commonalities and personalities of models in this new scenario and propose a novel approach named Federated Disentanglement Learning (FDISL). Due to significant variations in the data distribution across different clients, aggregating all parameters of clients’ models could cause the global model to skew in unpredictable directions. Therefore, we combine the essence of DG and the characteristics of FL to design the model from the perspective of disentanglement. Specifically, we divide the model into two parts: one extracts domain-invariant classification features and participates in global aggregation, while the other extracts domain-specific style features and updates locally only. To disentangle domain-invariant and domain-specific features without privacy leakage, FDISL utilizes a tiny unlabeled dataset to perform style transfer on each client’s local data, aligning domain-invariant features across styles and segregating domain-specific ones. To further boost performance, we employ distinct normalization strategies, mutual information minimization for enhanced decoupling, and contrastive learning to reinforce both category information in domain-invariant features and style information in domain-specific ones. Extensive experiments on several DG benchmarks demonstrate that FDISL achieves performance comparable to centralized DG methods and outperforms existing FedDG methods.
The rapid expansion of IoT and IIoT ecosystems has intensified security vulnerabilities, exposing interconnected devices to sophisticated and high-frequency cyberattacks. Traditional intrusion detection systems (IDS) often struggle with limited adaptability, high false alarms, and insufficient resilience against adversarial manipulation. To address these challenges, this paper proposes AI-QIDS, a comprehensive Adaptive AI-driven and Quantum-Resilient Intrusion Detection System that integrates deep learning, reinforcement learning, and post-quantum cryptographic protection. The intrusion detection engine is based on a classical hybrid CNN–BiLSTM architecture integrated with a Deep Q-Network (DQN) for adaptive threat detection, whereas quantum resilience is achieved through the incorporation of Kyber-768 lattice-based cryptography and QKD-assisted secure communication rather than quantum machine learning. Communication between IDS nodes is secured using Kyber-768 lattice-based encryption and QKD-derived symmetric keys. Extensive experiments on the Edge-IIoTset, comprising over 1.3 million normal and 15 attack classes, demonstrate that AI-QIDS achieves 99.36% accuracy, 99.25% precision, 99.42% recall, and 99.50% AUC, outperforming classical CNN, BiLSTM, DQN, and hybrid baselines. The system achieves 94.8% adversarial robustness against FGSM and PGD attacks, while introducing only a modest 8% latency overhead due to quantum-secure synchronization. A formal stability proof establishes bounded decision deviation under adversarial and quantum-induced perturbations. Ablation studies, complexity analysis, and statistical significance tests further validate the efficiency, scalability, and reliability of the proposed approach. Overall, AI-QIDS offers a future-proof and quantum-resilient intrusion detection paradigm suitable for large-scale IoT/IIoT deployments and next-generation cyber-physical systems.
Nowadays, swarms of drones are gaining a lot of interest for critical missions like search and rescue. They can be composed of several types of drones, with potentially different processing capabilities to successfully perform various computing tasks. In particular, some may use heterogeneous computing platforms, while others may rely on homogeneous parallel ones.Wireless communications used in the drone swarm are difficult to secure, due to their broadcasting properties and distributed aspects. In addition to cryptographic countermeasures, network Intrusion Detection Systems (IDSs) can act as a second layer of security to detect threats. To do so, they analyze network traffic to detect malicious behaviors, using resource-hungry Machine Learning (ML) models.Due to changing conditions during the swarm mission, it is necessary to embed several IDS implementations to adapt to these changes. This paper proposes two contributions. The first is IDS-DEEP, a strategy to select the best IDS for drones with a heterogeneous embedded platform. This strategy can be decomposed into two phases: (1) an offline phase to embed a subset of IDS implementations that cover a maximum of options, and (2) an online phase to select the best IDS implementation depending on the mission conditions. The second contribution is an extension of IDS-DEEP to support homogeneous (multi-core) parallel platforms, by leveraging on several cores to decrease latency of IDS model execution.Both strategies have been extensively validated experimentally using characterization results from several heterogeneous platforms (CPU, GPU, and FPGA), as well as a homogeneous parallel platform based on PULP. These contributions pave the way for safer and more efficient deployment of drone swarms in complex environments.
