This paper presents a hardware-efficient deep neural network (DNN), optimized through hardware-aware neural architecture search (HW-NAS); the DNN supports the classification of session-level encrypted traffic on resource-constrained Internet of Things (IoT) and edge devices. Thanks to HW-NAS, a 1D convolutional neural network (CNN) is tailored on the ISCX VPN-nonVPN dataset to meet strict memory and computational limits while achieving robust performance. The optimized model attains an accuracy of 96.60% with just 88.26K parameters, 10.08M floating-point operations (FLOPs), and a maximum tensor size of 20.12K. Compared to state-of-the-art (SOTA) models, it achieves reductions of up to 444-fold, 312-fold, and 15-fold in these metrics, respectively, significantly minimizing memory footprint and runtime requirements. The model also demonstrates versatility, achieving up to 99.86% across multiple VPN and traffic classification (TC) tasks; it further generalizes to external benchmarks with up to 99.98% accuracy on USTC-TFC and QUIC NetFlow. In addition, an in-depth approach to header-level preprocessing strategies confirms that the optimized model can provide notable performance across a wide range of configurations, even in scenarios with stricter privacy considerations. Likewise, a reduction in the length of sessions of up to 75% yields significant improvements in efficiency, while maintaining high accuracy with only a negligible drop of 1-2%. However, the importance of careful preprocessing and session length selection in the classification of raw traffic data is still present, as improper settings or aggressive reductions can bring about a 7% reduction in overall accuracy. The quantized architecture was deployed on STM32 microcontrollers and evaluated across input sizes; results confirm that the efficiency gains from shorter sessions translate to practical, low-latency embedded inference. These findings demonstrate the method’s practicality for encrypted traffic analysis in constrained IoT networks.

Smart sensors with extreme-edge data processing capabilities allow to optimize the energy budget of current Structural Health Monitoring (SHM) systems. Moving inference and Deep Neural Networks (DNNs) directly to the sensing unit can bypass the constraints due to data transmission, but it has to be properly balanced against the expenditure due to data sampling, which might easily become the main source of energy consumption when the duty cycling is not chosen appropriately. Finding the optimal sampling time and frequency is a non trivial task because it depends on the peculiarities of the target facility, the surrounding environment, and the highly non-linear behavior of DNNs. This paper presents a fully decentralized scheme for vibration damage detection where each sensor can predict the health status of the structure; a tiny Convolutional Neural Network (CNN) hosted on the node is entitled to autonomously process the measurements and to complete the inference. As a further contribution, an empirical study proves the major role played by the sampling parameters on the performance of the networks. Experiments on the Z24 bridge benchmark show that the small-size CNN can achieve classification scores comparable with state-of-the-art models while i) avoiding data transmission and ii) improving the energy budget if the optimal sensing parameters are chosen. These findings are obtained after deploying the sought CNN model on a wireless accelerometer sensor based on an STM32L496 processor. Results show that, by properly tuning the sampling parameters (length of the acquisition window and sampling frequency), one can reach the best trade-off between classification performances (accuracy above 96%) and energy consumption (less than 40 mJ/h).

In this paper, we present a practical deep learning (DL) approach for energy-efficient traffic classification (TC) on resource-limited microcontrollers, which are widely used in IoT-based smart systems and communication networks. Our objective is to balance accuracy, computational efficiency, and real-world deployability. To that end, we develop a lightweight 1D-CNN, optimized via hardware-aware neural architecture search (HW-NAS), which achieves $\mathbf{9 6. 5 9 \%}$ accuracy on the ISCX VPN-NonVPN dataset with only 88.26 K parameters, a 20.12 K maximum tensor size, and 10.08 M floating-point operations (FLOPs). Moreover, it generalizes across various TC tasks, with accuracies ranging from $94 \%$ to $99 \%$. To enable deployment, the model is quantized to INT8, suffering only a marginal $\mathbf{1} \boldsymbol{-} \mathbf{2} \boldsymbol{\%}$ accuracy drop relative to its Float32 counterpart. We evaluate real-world inference performance on two microcontrollers: the high-performance STM32F746G-DISCO and the cost-sensitive Nucleo-F401RE. The deployed model achieves inference latencies of 31.43 ms and 115.40 ms, with energy consumption of 7.86 mJ and 29.10 mJ per inference, respectively. These results demonstrate the feasibility of on-device encrypted traffic analysis, paving the way for scalable, low-power IoT security solutions.

Traffic classification (TC) plays a critical role in cybersecurity, particularly in IoT and embedded contexts, where inspection must often occur locally under tight hardware constraints. We use hardware-aware neural architecture search (HW-NAS) to derive lightweight TC models that are accurate, efficient, and deployable on edge platforms. Two input formats are considered: a flattened byte sequence and a 2D packetwise time series; we examine how input structure affects adversarial vulnerability when using resource-constrained models. Robustness is assessed against white-box attacks, specifically Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD). On USTC-TFC2016, both HW-NAS models achieve over $99 \%$ clean-data accuracy while remaining within 65 k parameters and 2 M FLOPs. Yet under perturbations of strength 0.1, their robustness diverges: the flat model retains over $\mathbf{8 5 \%}$ accuracy, while the time-series variant drops below 35%. Adversarial finetuning delivers robust gains, with flat-input accuracy exceeding $96 \%$ and the time-series variant recovering over 60 percentage points in robustness, all without compromising efficiency. The results underscore how input structure influences adversarial vulnerability, and show that even compact, resource-efficient models can attain strong robustness, supporting their practical deployment in secure edge-based TC.

This paper presents an automatic method for the design of Neural Networks (NNs) at the edge, enabling Machine Learning (ML) access even in privacy-sensitive Internet of Things (IoT) applications. The proposed method runs on IoT gateways and designs NNs for connected sensor nodes without sharing the collected data outside the local network, keeping the data in the site of collection. This approach has the potential to enable ML for Healthcare Internet of Things (HIoT) and Industrial Internet of Things (IIoT), designing hardware-friendly and custom NNs at the edge for personalized healthcare and advanced industrial services such as quality control, predictive maintenance, or fault diagnosis. By preventing data from being disclosed to cloud services, this method safeguards sensitive information, including industrial secrets and personal data. The outcomes of a thorough experimental session confirm that –on the Visual Wake Words dataset– the proposed approach can achieve state-of-the-art results by exploiting a search procedure that runs in less than 10 hours on the Raspberry Pi Zero 2.

Internet-of-Things (IoT) is a key enabler for the transition to the Automatic Structural Health Monitoring (ASHM) of technical facilities, thanks to the seamless flow of data from a multitude of always connected devices. Current IoT-ASHM installations, however, face the double challenge to ensure high accuracy while meeting the requirement of minimal energy consumption. The paper tackles these issues from a deep-learning perspective, and describes an IoT-enabled monitoring approach based on a distributed end-to-end deep neural network (DNN). The architecture supports both data compression and damage detection. A low-end microcontroller hosts a specific local DNN; a hardware-aware neural-architecture search strategy rules network optimization, in order to satisfy the resource constraints set by low-end computing devices. The features extracted from data feed an aggregating unit, which includes a stacked global classification layer for full-scale damage detection. After proper quantization, the designed models are eventually deployed on a wireless accelerometer sensor. Finally, a cost-benefit analysis evaluates the system’s impact on the sensor energy autonomy. Experiments on a well-known dataset proved that the proposed solution could achieve state-of-the-art classification scores (all metrics above 98.4%) with a minimal transmission cost (less than 53 B on average); as compared with conventional approaches, the described strategy yielded a reduction of three orders of magnitude in energy consumption.

Compressed sensing (CS) for sensor-near vibration diagnostics represents a suitable approach for the design of network-efficient structural health monitoring systems. This article presents a solution for vibration analysis based on deep neural networks (DNNs) trained on compressed data. The envisioned maintenance system consists of a network of sensing nodes orchestrated by a very constrained centralizing unit. The latter is equipped with a microcontroller unit (MCU) that predicts the health state using the aggregated information. As a major contribution, the DNN architectures are generated automatically from the data through a procedure inspired by hardware-aware (HW) neural architecture search (NAS), called as HW-NAS-CS, which is uniquely refined with additional constraints that consider both the peculiarities of CS parameters and the limitation of embedded devices. The proposed approach has been validated using two real-world SHM datasets for vibration damage identification and eventually deployed on a low-end computing platform (the STM32L5 MCU). Results demonstrate that DNNs combined with adapted CS schemes can attain classification scores always above 90% even in case of very huge compression levels (higher than 64x): these performances significantly improve the ones attained by state-of-the-art approaches in the field, with the utmost advantage of being portable on embedded devices.

The availability of new datasets and deep learning techniques have led to a surge of effort directed towards the creation of new models that can exploit the large amount of data. However, little attention has been given to the development of models that are not only accurate, but also suitable for user-specific use or geared towards resource-constrained devices. Fine-tuning deep models on edge devices is impractical and, often, user customization stands on the sub-optimal feature-extractor/classifier paradigm. Here, we propose a method to fully utilize the intermediate outputs of the popular large pre-trained models in natural language processing when used as frozen feature extractors, and further close the gap between their fine-tuning and more computationally efficient solutions. We reach this goal exploiting the concept of software-hardware co-design and propose a methodical procedure, inspired by Neural Architecture Search, to select the most desirable model taking into consideration application constraints.

Abstract Variable energy constraints affect the implementations of neural networks on battery-operated embedded systems. This paper describes a learning algorithm for randomization-based neural networks with hard-limit activation functions. The approach adopts a novel cost function that balances accuracy and network complexity during training. From an energy-specific perspective, the new learning strategy allows to adjust, dynamically and in real time, the number of operations during the network’s forward phase. The proposed learning scheme leads to efficient predictors supported by digital architectures. The resulting digital architecture can switch to approximate computing at run time, in compliance with the available energy budget. Experiments on 10 real-world prediction testbeds confirmed the effectiveness of the learning scheme. Additional tests on limited-resource devices supported the implementation efficiency of the overall design approach.

In the research area of human-robot interactions, the automatic estimation of the mass of a container manipulated by a person leveraging only visual information is a challenging task. The main challenges consist of occlusions, different filling materials and lighting conditions. The mass of an object constitutes key information for the robot to correctly regulate the force required to grasp the container. We propose a single RGB-D camera-based method to locate a manipulated container and estimate its empty mass i.e., independently of the presence of the content. The method first automatically selects a number of candidate containers based on the distance with the fixed frontal view, then averages the mass predictions of a lightweight model to provide the final estimation. Results on the CORSMAL Containers Manipulation dataset show that the proposed method estimates empty container mass obtaining a score of 71.08% under different lighting or filling conditions.

Abstract Random-based learning paradigms exhibit efficient training algorithms and remarkable generalization performances. However, the computational cost of the training procedure scales with the cube of the number of hidden neurons. The paper presents a novel training procedure for random-based neural networks, which combines ensemble techniques and dropout regularization. This limits the computational complexity of the training phase without affecting classification performance significantly; the method best fits Internet of Things (IoT) applications. In the training algorithm, one first generates a pool of random neurons; then, an ensemble of independent sub-networks (each including a fraction of the original pool) is trained; finally, the sub-networks are integrated into one classifier. The experimental validation compared the proposed approach with state-of-the-art solutions, by taking into account both generalization performance and computational complexity. To verify the effectiveness in IoT applications, the training procedures were deployed on a pair of commercially available embedded devices. The results showed that the proposed approach overall improved accuracy, with a minor degradation in performance in a few cases. When considering embedded implementations as compared with conventional architectures, the speedup of the proposed method scored up to 20× in IoT devices.

Providing user customized experience is one of the main goals for present-day electronic smart devices. Image polarity detection plays a crucial role in understanding users' preferences due to the fact that information is massively represented by means of pictures. State-of-the-art frameworks are based on deep learning networks and continue evolving adding sophisticated structures to enhance generalization performances of the inference systems. Recent works proved that image analysis can be enhanced exploiting the information about salient regions. However, better performances are obtained at the cost of a higher computational load. This paper presents a hardware-friendly deep learning framework for image polarity detectors based on salient regions of an image. Experimental results show the reliable performances of the proposed solution on real-world data.

Law Enforcement Agencies cover a crucial role in the analysis of open data and need effective techniques to filter troublesome information. In a real scenario, Law Enforcement Agencies analyze Social Networks, i.e. Twitter, monitoring events and profiling accounts. Unfortunately, between the huge amount of internet users, there are people that use microblogs for harassing other people or spreading malicious contents. Users' classification and spammers' identification is a useful technique for relieve Twitter traffic from uninformative content. This work proposes a framework that exploits a non-uniform feature sampling inside a gray box Machine Learning System, using a variant of the Random Forests Algorithm to identify spammers inside Twitter traffic. Experiments are made on a popular Twitter dataset and on a new dataset of Twitter users. The new provided Twitter dataset is made up of users labeled as spammers or legitimate users, described by 54 features. Experimental results demonstrate the effectiveness of enriched feature sampling method.