The rise of online platforms exacerbated the spread of hate speech, demanding scalable and effective detection. However, the accuracy of hate speech detection systems heavily relies on human-labeled data, which is inherently susceptible to biases. While previous work has examined the issue, the interplay between the characteristics of the annotator and those of the target of the hate are still unexplored. We fill this gap by leveraging an extensive dataset with rich socio-demographic information of both annotators and targets, uncovering how human biases manifest in relation to the target's attributes. Our analysis surfaces the presence of widespread biases, which we quantitatively describe and characterize based on their intensity and prevalence, revealing marked differences. Furthermore, we compare human biases with those exhibited by persona-based LLMs. Our findings indicate that while persona-based LLMs do exhibit biases, these differ significantly from those of human annotators. Overall, our work offers new and nuanced results on human biases in hate speech annotations, as well as fresh insights into the design of AI-driven hate speech detection systems.

Deepfakes are a type of synthetic media created using artificial intelligence, specifically deep learning algorithms. This technology can for example superimpose faces and voices onto videos, creating hyper-realistic but artificial representations. Deepfakes pose significant risks regarding misinformation and fake news, because they can spread false information by depicting public figures saying or doing things they never did, undermining public trust. In this paper, we propose a method that leverages BLS signatures (Boneh, Lynn, and Shacham 2004) to implement signatures that remain valid after image cropping, but are invalidated in all the other types of manipulation, including deepfake creation. Our approach does not require who crops the image to know the signature private key or to be trusted in general, and it is ${\mathcal{O}}(1)$ in terms of signature size, making it a practical solution for scenarios where images are disseminated through web servers and cropping is the primary transformation. Finally, we adapted the signature scheme for the JPEG standard, and we experimentally tested the size of a signed image.

Recently, object counting has shifted towards classagnostic counting (CAC), which counts instances of arbitrary object classes never seen during model training. With advancements in robust vision-and-language foundation models, there is a growing interest in prompt-based CAC, where object categories are specified using natural language. However, we identify significant limitations in current benchmarks for evaluating this task, which hinder both accurate assessment and the development of more effective solutions. Specifically, we argue that the current evaluation protocols do not measure the ability of the model to understand which object has to be counted. This is due to two main factors: (i) the shortcomings of CAC datasets, which primarily consist of images containing objects from a single class, and (ii) the limitations of current counting performance evaluators, which are based on traditional class-specific counting and focus solely on counting errors. To fill this gap, we introduce the Prompt-Aware Counting (PrACo) benchmark. It comprises two targeted tests coupled with evaluation metrics specifically designed to quantitatively measure the robustness and trustworthiness of existing prompt-based CAC models. We evaluate state-of-the-art methods and demonstrate that, although some achieve impressive results on standard class-specific counting metrics, they exhibit a significant deficiency in understanding the input prompt, indicating the need for more careful training procedures or revised designs. The code for reproducing our results is available at https://github.com/ciampluca/PrACo.

AI-generated counterspeech offers a promising and scalable strategy to curb online toxicity through direct replies that promote civil discourse. However, current counterspeech is one-size-fits-all, lacking adaptation to the moderation context and the users involved. We propose and evaluate multiple strategies for generating tailored counterspeech that is adapted to the moderation context and personalized for the moderated user. We instruct a LLaMA2-13B model to generate counterspeech, experimenting with various configurations based on different contextual information and fine-tuning strategies. We identify the configurations that generate persuasive counterspeech through a combination of quantitative indicators and human evaluations collected via a pre-registered mixed-design crowdsourcing experiment. Results show that contextualized counterspeech can significantly outperform state-of-the-art generic counterspeech in adequacy and persuasiveness, without compromising other characteristics. Our findings also reveal a poor correlation between quantitative indicators and human evaluations, suggesting that these methods assess different aspects and highlighting the need for nuanced evaluation methodologies. The effectiveness of contextualized AI-generated counterspeech and the divergence between human and algorithmic evaluations underscore the importance of increased human-AI collaboration in content moderation.

The COVID-19 pandemic has considerably shifted the focus of scientific research, speeding up the process of digitizing medical monitoring. Wearable technology is already widely used in medical research, as it has the potential to monitor the user’s physical activity in daily life. This study aims to explore in-home collected wearable-derived signals for frailty status assessment. A sample of 35 subjects aged 70+, autonomous in basic activities of daily living and cognitively intact, was collected. After being clinically assessed for frailty according to Fried’s phenotype, participants wore a wrist device equipped with inertial motion sensors for 24 h, during which they led their usual life in their homes. Signal-derived traces were split into 10-second segments and labeled classified as gaits, other motor activities, or rests. Gait and other motor activity segments were used to calculate the Subject Activity Level (SAL), an index to quantify how users were active throughout the day. The SAL index was then combined with gait-derived features to design a novel frailty status assessment algorithm. In particular, subjects were classified as robust or non-robust, a category that includes both Fried’s frail and pre-frail phenotypes. For some users, activity levels alone enabled accurate frailty assessment, whereas, for others, a Gaussian Naive Bayes classifier based on the gait-derived features was required to assess frailty status. Overall, the proposed method showed extremely promising results, allowing discrimination of robust and non-robust subjects with an overall 91% accuracy, stemming from 95% sensitivity and 88% specificity. This study demonstrates the potential of unobtrusive, wearable devices in objectively assessing frailty through unsupervised monitoring in real-world settings.

Automatic people counting from images has recently drawn attention for urban monitoring in modern Smart Cities due to the ubiquity of surveillance camera networks. Current computer vision techniques rely on deep learning-based algorithms that estimate pedestrian densities in still, individual images. Only a bunch of works take advantage of temporal consistency in video sequences. In this work, we propose a spatio-temporal attentive neural network to estimate the number of pedestrians from surveillance videos. By taking advantage of the temporal correlation between consecutive frames, we lowered state-of-the-art count error by 5% and localization error by 7.5% on the widely-used FDST benchmark.

Wearable sensors potentially enable monitoring the user's physical activity in daily life. Therefore, they are particularly appealing for the evaluation of older subjects in their environment, to capture early signs of frailty and mobility-related problems. This study explores the use of body-worn accelerometers for automated assessment of frailty during walking activity. Experiments involved 34 volunteers aged 70+, who were initially screened by geriatricians for the presence of frailty according to Fried's criteria. After screening, the volunteers were asked to walk 60 m at preferred speed, while wearing two accelerometers, one positioned on the lower back and the other on the wrist. Sensor-derived signals were analyzed independently to compare the ability of the two signals (wrist vs. lower back) in frailty status assessment. A gait detection technique was applied to identify segments made of four gait cycles. These segments were then used as input to compute 25 features in time and time-frequency domains, the latter by means of the Wavelet Transform. Finally, five machine learning models were trained and evaluated to classify subjects as robust or non-robust (i.e., pre-frail or frail). Gaussian naive Bayes applied to the features derived from the wrist sensor signal identified non-robust subjects with 91% sensitivity and 82% specificity, compared to 87% sensitivity and 64% specificity achieved with the lower back sensor. Results demonstrate that a wrist-worn accelerometer provides valuable information for the recognition of frailty in older adults, and could represent an effective tool to enable automated and unobtrusive assessment of frailty.

Recent advancements in the field of smart wearable sensors provide the opportunity of continuous analysis of user's movements, which enables the assessment of clinical conditions like frailty. This study explores the use of Continuous Wavelet Transform in combination with sensor-derived gait parameters for frailty status assessment. A total of 34 volunteers aged 70+ were initially screened by geriatricians for the presence of frailty according to Fried's criteria. After screening, participants were asked to perform a 60 m walk test at preferred pace, while wearing an accelerometer on the wrist. A gait detection technique was applied to the sensor-derived signal, in order to identify segments made of four gait cycles. Continuous Wavelet Transform was applied to obtain time-frequency domain representations, which were subsequently used in a band-based feature extraction phase. Here, the most significant band-based features for frailty status assessment were identified by means of ANOVA and statistical t-test. Finally, a Random Forest for each frequency band was trained and tested for classifying subjects as robust or nonrobust (i.e., pre-frail or frail). Results from both the statistical analysis and machine learning show that features extracted from [1.5, 2.5]Hz frequency band can provide valuable information for recognizing frailty in older adults. This information may help achieve continuous assessment of frailty in older adults with a wrist-worn device.

Wrist-worn devices, such as smartwatches and smart bands, have brought about the unprecedented opportunity to continuously monitor gait during daily routines. However, the use of a single wrist-worn unit for gait analysis is challenging for a variety of reasons. Indeed, the signal collected at the user's wrist is subject to a significant “noise” with respect to other body positions (e.g. waist), mainly due to the arm swing while walking and other unpredictable hand movements. The aim of this paper is to investigate the design and evaluation of a lightweight and reliable gait detection technique for wrist-worn devices. To this end, the proposed method creates a personalized model of the user's gait patterns. The model is created through an automatic training phase, which requires the temporary use of an additional device (smartphone) to gather true gait segments. After, anomaly detection is used to distinguish gait from other activities. Gait data from 20 volunteers have been collected to test and evaluate the proposed technique. Volunteers were asked to walk at different pace, with their normal arm swing or placing the hand inside of a pocket. Results show that the proposed method can reliably distinguish gait from spurious hand movements.

Wearable devices can gather sensitive information about their users. For this reason, automated authentication and identification techniques are increasingly adopted to ensure security and privacy. Furthermore, identification can be used to automatically customize operations according to the needs of the current user. A gait-based identification method that can be executed in real time on devices with limited resources is here presented. The method exploits a wearable accelerometer to continuously analyze the user's gait pattern and perform identification. Experiments were conducted with 10 volunteers, who carried the device in a trouser pocket and followed their daily routine without predefined constraints. In total, ∼98 hours of acceleration traces were collected in uncontrolled environment, including 3073 gait segments. User identification results show a recognition rate ranging from 95% to 100%, depending on the mode of operation. It is demonstrated that the method can be executed on a standalone device with <8 KB of RAM. In addition, the energy consumption is evaluated and compared with an architecture that requires the presence of an external computing unit. Results show that the proposed solution significantly improves the lifetime of the device (approximately +70% for the considered platform), hence fostering user acceptance.

Every individual has a distinctive way of walking. For this reason gait can be a key element of biometric techniques aimed at authenticating and/or identifying the user of a wearable device. This paper presents a lightweight method that uses the acceleration collected at the user's wrist for authentication purposes. The user's typical gait pattern is learned during the first period of use, then detection of anomalies in a set of acceleration-based features is used to understand if a new user, a possible impostor or a thief, is wearing the device. The method has been successfully evaluated with 15 volunteers, showing an Equal Error Rate of 2.9%. These results suggest that gait-based authentication with a wrist-worn device can be carried out with high accuracy levels. A comparison with a similar method executed on a pocket-worn device is also included.

Similar to fingerprint and iris pattern, everyone's gait is unique, and gait has been proposed as a biometric feature for security applications. This paper presents a lightweight accelerometer-based technique for user authentication on smart wearable devices. Designed as an unsupervised classification approach, the proposed authentication technique can learn the user's gait pattern automatically when the user first starts wearing the device. Anomaly detection is then used to verify the device owner. The technique has been evaluated both in controlled and uncontrolled environments, with 20 and 6 healthy volunteers respectively. The Equal Error Rate (EER) in the controlled environments ranged from 5.7% (waist-mounted sensor) to 8.0% (trouser pocket). In the uncontrolled experiment, the device was put in the subject's trouser pocket, and the results were similar to the respective supervised experiment (EER=9.7%).

Social sensing is based on the idea that communities or groups of people can provide a set of information similar to those obtainable from a sensor network. Emergency management is a candidate field of application for social sensing. In this work we describe the design, implementation and deployment of a decision support system for the detection and the damage assessment of earthquakes in Italy. Our system exploits the messages shared in real-time on Twitter, one of the most popular social networks in the world. Data mining and natural language processing techniques are employed to select meaningful and comprehensive sets of tweets. We then apply a burst detection algorithm in order to promptly identify outbreaking seismic events. Detected events are automatically broadcasted by our system via a dedicated Twitter account and by email notifications. In addition, we mine the content of the messages associated to an event to discover knowledge on its consequences. Finally we compare our results with official data provided by the National Institute of Geophysics and Volcanology (INGV), the authority responsible for monitoring seismic events in Italy. The INGV network detects shaking levels produced by the earthquake, but can only model the damage scenario by using empirical relationships. This scenario can be greatly improved with direct information site by site. Results show that the system has a great ability to detect events of a magnitude in the region of 3.5, with relatively low occurrences of false positives. Earthquake detection mostly occurs within seconds of the event and far earlier than the notifications shared by INGV or by other official channels. Thus, we are able to alert interested parties promptly. Information discovered by our system can be extremely useful to all the government agencies interested in mitigating the impact of earthquakes, as well as the news agencies looking for fresh information to publish.

The most advanced solutions that are currently adopted in ports and terminals use technologies based on radio frequency identification (RFID) and the Global Positioning System (GPS) to identify and localize shipping containers in the yard. Nevertheless, because of the limitations of these solutions, the position of containers is still affected by errors, and it cannot be determined in real time. In this paper, a nonconventional approach is presented: Each container is equipped with nodes that use wireless communication to detect neighbor containers and to send proximity information to a base station. At the base station, geometrical constraints and proximity data are combined to determine the positions of containers. Missing information due to faulty nodes is tolerated by modeling geometrical constraints as an integer linear programming problem. Numerical simulations show that most of the containers can be localized, even when the number of nodes that are affected by faults is on the order of 30%.

Run-time monitoring of energy consumption in wireless sensor networks is a necessary step for the production of energy efficient applications. The demo will show a software system that helps the developer to profile applications based on TinyOS in terms of energy consumption.

Falls are a major cause of hospitalization and injury-related deaths among the elderly population. The detrimental effects of falls, as well as the negative impact on health services costs, have led to a great interest on fall detection systems by the health-care industry. The most promising approaches are those based on a wearable device that monitors the movements of the patient, recognizes a fall and triggers an alarm. Unfortunately such techniques suffer from the problem of false alarms: some activities of daily living are erroneously reported as falls, thus reducing the confidence of the user. This paper presents a novel approach for improving the detection accuracy which is based on the idea of identifying specific movement patterns into the acceleration data. Using a single accelerometer, our system can recognize these patterns and use them to distinguish activities of daily living from real falls; thus the number of false alarms is reduced.

In order to obtain maximum energetic efficiency, wireless sensor networks must be able to tailor their mode of operation at every level of the software infrastructure. At the MAC level, adaptability can greatly improve the performance of the system by tuning the parameters of operation of the protocol on the base of the communication pattern of the application. This paper presents an adaptive, contention-based MAC protocol where the duty cycle of the transceiver is dynamically changed to match the traffic rate. Adaptation is achieved by using local information and without generating additional messages. Coordination among nodes is based on passive dissemination of nodes' low power listening modes.

Wireless sensor networks are moving from academia to real world scenarios. This will involve, in the near future, the design and production of hardware platforms characterized by low-cost and small form factor. As a consequence, the amount of resources available on a single node, i.e. computing power, storage, and energy, will be even more constrained than today. This paper faces the problem of storing and executing an application that exceeds the memory resources available on a single node. The proposed solution is based on the idea of partitioning the application code into a number of opportunistically cooperating modules. Each node contributes to the execution of the original application by running a subset of the application tasks and providing service to the neighboring nodes.

Abstract Code mobility is recognized as a promising design technique, able to improve flexibility, adaptability and bandwidth utilization in mobile computing applications. To promote and facilitate its use, researchers argue that code mobility should be made available to programmers in combination with, and not as an alternative to, more traditional programming models. This paper describes the design and implementation of the MobileRMI toolkit which, unlike agent‐based systems, enables mobility‐based programming within a widely accepted middleware platform, Java Remote Method Invocation (RMI). Our toolkit provides a set of mobility primitives that allow programmers to create, clone and move remote objects across a network. To preserve location transparency we implemented a novel, efficient scheme for automatically updating remote references by exploiting the distributed garbage collector. Programming examples are given and a case study where an adaptive application uses logical mobility to minimize communication over a mobile ad hoc network is presented. Experience from using MobileRMI confirmed the benefit of designing both static and mobile applications within the same programming framework. Copyright © 2005 John Wiley & Sons, Ltd.

Network emulation provides the capability of evaluating distributed applications on a stand-alone system. Applications can be exposed to adverse and repeatable network conditions without requiring complex testbeds. This paper describes the design and implementation of a portable and object-oriented network emulator targeted to the development and test of Java-based Internet applications. The emulator is based on instrumented sockets, say EmuSockets, able to emulate the behavior of links with a given bandwidth and communication delay. The emulator is organized modularly, so that it is possible to plug-in user-defined classes for bandwidth and delay figures generation. Carrying out experiments with EmuSockets is as simple as running the tested application code on a single host.