Network intrusion detection systems (NIDS) are critical for safeguarding modern organizations against an increasing number of different cyber threats. Traditional NIDS approaches, whether signature-based or anomaly-based, exhibit limitations in detecting novel or complex attacks and often struggle with adaptability in dynamic network environments. This paper introduces LLM-NIRED (Large Language Model-based Network Intrusion Robust Efficient Detection), a novel network intrusion detection system (NIDS) leveraging advanced Large Language Models (LLMs) that aims to provide a flexible, memory efficient solution that is robust to previously unseen networks and suitable for running on edge devices. In this context, we define ‘robust’ as the model’s demonstrated capacity to generalize effectively across unseen network topologies and novel traffic distributions, a common failing point for traditional deep learning-based NIDS. By treating network traffic as sequences of tokens, our fine-tuned model identifies complex attack patterns. While rich anomaly-based profiles constructed from the same datasets can also detect similar traffic behaviors, our LLM-based approach natively captures sequential dependencies, with accuracy approaching 99% on zero-overlap external datasets. It outperforms state-of-the-art deep learning and BERT-based NIDS in generalization, while exhibiting better memory efficiency and inference times comparable to existing transformer-based approaches. Furthermore, we explore the feasibility of deploying a lightweight variant of LLM-NIRED (TinyLLM) on resource-constrained edge devices, showcasing its potential for on-edge analysis with minimal computational overhead.

Recent developments in smart assistants and smart home automation are lately attracting the interest and curiosity of consumers and researchers. Speech enabled virtual assistants (often named smart speakers) offer a wide variety of network-oriented services and, in some cases, can connect to smart environments, thus enhancing them with new and effective user interfaces. However, such devices also reveal new needs and some weaknesses. In particular, they represent faceless and blind assistants, unable to show a face, and therefore an emotion, and unable to `see' the user. As a consequence, the interaction is impaired and, in some cases, ineffective. Moreover, most of those devices heavily rely on cloud-based services, thus transmitting potentially sensitive data to remote servers. To overcome such issues, in this paper we combine some of the most advanced techniques in computer vision, deep learning, speech generation and recognition, and artificial intelligence, into a virtual assistant architecture for smart home automation systems. The proposed assistant is effective and resource-efficient, interactive and customizable, and the realized prototype runs on a low-cost, small-sized, Raspberry PI 3 device. For testing purposes, the system was integrated with an open source home automation environment and ran for several days, while people were encouraged to interact with it, and proved to be accurate, reliable and appealing.

Face recognition represents an appealing solution for biometrics-based unobtrusive and flexible person authentication. However, most solutions proposed in the literature suffer from some significant drawbacks, such as high computational complexity, the need for a centralized biometrics database (which is not desirable due to widespread international provisions discouraging collections of sensitive personal data), and limited scalability on a large number of enrolled subjects. We propose a novel person authentication solution based on a cascade of face recognition and pattern matching algorithms that not only provides high reliability and robustness against impostors but also stores in a personal radio frequency identification (RFID) tag all the needed individual biometrics information of the user, who therefore always remains in control, and has the exclusive availability, of such sensitive data. This paper describes the proposed approach, called RFaceID, and discusses its performance in terms of the ratio between false acceptance rate and false rejection rate and in terms of authentication time when applied to the VidTIMIT, Extended Yale B, and Mobile Biometry (MOBIO) widely adopted face databases.

This paper proposes the Multilevel Gaussian Mixture Model (ML_GMM), a novel algorithm for online, unsupervised modeling of complex motion trajectories which allows for fast and reliable abnormal trajectory detection in video surveillance. The presented approach does not directly deal with target tracking or attribution of an observation to either one or another trajectories belonging to different targets. Instead, it receives as an input a sequence of observations and considers them as produced by a reliable, yet possibly noisy, tracker. As an output, it then produces a labeling of the input sequence, as a "normal" or "abnormal" whole trajectory. The normal and abnormal trajectory domains are learned in an unsupervised fashion during the training phase and can be updated while the classifier is running, thanks to its reduced computational complexity. The novelty of the contribution is twofold. First, ML_GMM is capable of representing short-term, as well as long-term, transitions in a consistent way, thus allowing for accurate representation of long and complex trajectories and their reliable labeling as normal or abnormal. Second, it allows the learning process to be run in parallel with the discrimination process, in order to support adaptability of the model to evolving operating conditions. Experimental results show that ML_GMM outperforms a number of comparable approaches in the literature.

The Media Oriented Systems Transport (MOST) network, the current de-facto standard for in-vehicle multimedia and infotainment, does not provide support for the stream priority and serves bandwidth allocation requests based on their arrival order, irrespective of the stream importance. This work proposes a novel bandwidth allocation approach for multimedia streams on MOST150 networks that introduces the stream priority, which is used to reserve bandwidth for each stream according to its importance. The approach also exploits the ability for dynamically changing the bandwidth needs of multimedia streams through variable bitrate compression and is useful in MOST 150 networks supporting, together with multimedia streams, also bandwidth-greedy applications, such as Advanced Driver Assistance Systems.

Advanced Driver Assistance Systems involving cameras and the multimedia domain are both quite challenging for automotive network technologies, due to the bandwidth requirements and latency/jitter constraints they impose. Ethernet as in-car network is expected to breakthrough in both domains and the IEEE Audio Video Bridging (AVB) standard is a promising candidate. This paper reports preliminary performance assessments of AVB for ADAS, multimedia and infotainment, obtained through the simulation of realistic traffic patterns.

Vision-based, human-computer interaction technology is mature to substitute legacy physical devices such as mouse and joy stick for a number of different applications. Yet, some sort of mistrust prevents them from being used mainly because - in the past - the main focus has been on the computer vision side and not on usability, reliability and efficiency of the interface. This paper introduces a vision-based perceptual user interface (PUI) applied to radio-controlled (RC) mobile units. Its main effort is on shifting the research focus from basic computer vision and image processing issues to ease and naturalness of use, integrability and compatibility with the existing systems, portability and efficiency. We believe that current technologies can easily support vision-based PUIs and that PUIs are strongly needed by modern applications, but more effort is needed to merge the knowledge from HCI and computer vision communities to develop realistic and industrially appealing products. A proof-of-concept application is described and experimental results are provided and discussed.

When strong CPU power consumption constraints must be met, and high computation speed is mandatory (real-time processing), it can be preferable to adopt custom hardware for some computationally intensive image processing tasks. An alternative approach to the conventional ones is provided by the cellular neural network (CNN) paradigm. CNNs have been extensively used in image processing applications: in the past, we developed a still image segmentation technique based on an active contour obtained via single-layer CNNs. This technique suffered from sensitivity to noise as most of edge-based methods: noise may create meaningless false edges or determine "edge fragmentation". The aim of this paper is to reformulate the algorithm previously proposed in order to step-over the cited weakness. The new formulation is introduced and motivated and experimental results are presented. Finally, a competition-based approach for a parameterless version of the presented algorithm is proposed and discussed as an ongoing work.

In this paper a perceptual user interface (PUI) for video-surveillance environments is introduced. This system provides a tool for a video-surveillance control-room, and exploits a novel multimodal user interaction paradigm based on hand gesture and perceptual user interfaces. The proposed system, being simple and intuitive, is expected to be useful in the control of large and dynamic environments. To illustrate our work, we introduce a proof-of concept multimodal, bare-hand gesture-based application and discuss its implementation and the obtained experimental results.

In this paper we present a new real-time algorithm for object tracking by on-line video analysis. This algorithm is able to track an object even in presence of substantial perspective transformations and to detect changes in the object shape due to changes in its structure. The core of the system is a chain made of autonomous software agents (MOVels), which acts like an active contour on the sequence of frames.

In this paper we propose an edge-based segmentation algorithm built on a new type of active contour which is fast, has a low computational complexity and does not introduce unwanted smoothing on the retrieved contours. The contours are always returned as closed chains of points, resulting in a very useful base for subsequent shape representation techniques.

We have developed a fast method to segment 2-D grey-level images, obtaining closed-line, continuous, smooth and analytical representations of object contours. The results are very good, even in presence of noisy or low contrast images, and the shape of objects are retrieved with very good accuracy.