Human voice is a rich source of information that can reveal a range of sensitive personal attributes, such as age, gender, and country of origin. With advances in Artificial Intelligence (AI), especially in speech processing, these personal attributes can now be inferred on a scale with high accuracy, raising serious privacy concerns. In fact, the ability to extract demographic or identification information from voice data poses risks related to surveillance, profiling, and misuse of personal data, highlighting the urgent need for privacy-preserving solutions in voice-based AI systems. Therefore, this paper proposes Selective Attribute Masking (SAM) , a new model-agnostic framework that uses gradient-based adversarial perturbations to suppress the inference of specific speaker attributes from voice recordings, while preserving the accuracy of non-target attributes and maintaining the utility of Automatic Speech Recognition (ASR). Experimental results using CommonVoice dataset demonstrate that SAM achieves selective masking success rates of up to 74.5 % for age, 59.4 % for gender, and 54.6 % for accent–substantially outperforming baseline methods. At the same time, voice utility (that is, ASR) remains largely unaffected, with the word error rate increasing by less than 3 % absolute under moderate perturbations. These findings demonstrate the effectiveness of our proposed framework (SAM) in balancing privacy and utility in voice-based systems.

Large language models (LLMs) have transformed NLP, yet their integration with audio remains underex-plored-despite audio’s centrality to human communication. We introduce Falcon3-Audio, a family of Audio-Language Models (ALMs) built on instruction-tuned LLMs and Whisper encoders. Using a remarkably small amount of public audio data-less than 30 K hours (5 K unique)-Falcon3-Audio-7B matches the best reported performance among open-weight models on the MMAU benchmark, with a score of 64.14, matching R1-AQA, while distinguishing itself through superior data and parameter efficiency, single-stage training, and transparency. Notably, our smallest 1B model remains competitive with larger open models ranging from 2 B to 13 B parameters. Through extensive ablations, we find that common complexities-such as curriculum learning, multiple audio encoders, and intricate cross-attention connec-tors-are not required for strong performance, even compared to models trained on over $\mathbf{5 0 0 K}$ hours of data.

Multi-level Hierarchical Classification (MLHC) tackles the challenge of categorizing items within a complex, multi-layered class structure. However, traditional MLHC classifiers often rely on a backbone model with n independent output layers, which tend to ignore the hierarchical relationships between classes. This oversight can lead to inconsistent predictions that violate the underlying taxonomy. Leveraging Large Language Models (LLMs), we propose novel taxonomy-embedded transitional LLM-agnostic framework for multimodality classification. The cornerstone of this advancement is the ability of models to enforce consistency across hierarchical levels. Our evaluations on the MEP-3M dataset - a Multi-modal E-commerce Product dataset with various hierarchical levels- demonstrated a significant performance improvement compared to conventional LLMs structure.

Synthetic data has gained attention for training large language models, but poor-quality data can harm performance (see, e.g., Shumailov et al. (2023); Seddik et al. (2024)). A potential solution is data pruning, which retains only high-quality data based on a score function (human or machine feedback). Previous work Feng et al. (2024) analyzed models trained on synthetic data as sample size increases. We extend this by using random matrix theory to derive the performance of a binary classifier trained on a mix of real and pruned synthetic data in a high dimensional setting. Our findings identify conditions where synthetic data could improve performance, focusing on the quality of the generative model and verification strategy. We also show a smooth phase transition in synthetic label noise, contrasting with prior sharp behavior in infinite sample limits. Experiments with toy models and large language models validate our theoretical results.

Salem Lahlou, Abdalgader Abubaker, Hakim Hacid. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

In critical domains such as healthcare and law, accurately modelling the uncertainty of automatic computational models is essential. For instance, healthcare models must produce reliable estimates to guide human decision-making. However, modelling uncertainty remains challenging, particularly for models handling low-resource datasets and complex, domain-specific vocabulary. Most existing predictive models model point estimates rather than probability distributions, limiting our ability to quantify model uncertainty. This paper introduces a novel model, BAKER, designed to address these limitations. BAKER combines the strengths of Bayesian inference, known for its effectiveness in modelling uncertainty, and kernel methods, which excel at capturing complex data relationships. Incorporating kernel functions enhances model performance, particularly by reducing overfitting in data-limited scenarios. Our experimental analysis shows that BAKER significantly improves uncertainty reasoning compared to existing models.

Audio profiling is at the forefront of a technological breakthrough, offering rich insights into human behavior, emotions, physical attributes, and environmental contexts through detailed analysis of voice data. As we embrace an era where the integration of smart technologies equipped with the ability to capture sound is becoming ubiquitous, the capacity to accurately infer personal traits such as age, gender, height, weight, emotional state , personality, and even environmental contexts through voice analysis opens up vast opportunities across law enforcement, healthcare, social and commercial services, and entertainment. This emerging field promises to enhance our interaction with technology by not only understanding who we are but also by interpreting the world around us. However, the remarkable landscape is fraught with challenges, including data imbalances, the complexity of predictive models , and significant privacy concerns regarding the handling of sensitive paralinguistic information. This survey explores deep into the current landscape of audio profiling, examining the techniques and datasets in use, and showcasing its diverse applications while highlighting the need for advanced methodologies, enriched dataset development, and robust privacy preservation techniques.

This demonstration introduces SOREO, a system that explores the possibility of extending UAVs autonomy through machine learning. It brings a contribution to the following problem: Having a fleet of drones and a geographic area, how to learn the shortest paths between any point with regards to the base points for optimal and safe package delivery? Starting from a set of possible actions, a virtual design of a geographic location of interest, e.g., a city, and a reward value, SOREO is capable of learning not only how to prevent collisions with obstacles, e.g., walls and buildings, but also to find the shortest path between any two points, i.e., the base and the target. SOREO exploits based on the Q-learning algorithm.

Edge Machine Learning (Edge ML), which shifts computational intelligence from cloud-based systems to edge devices, is attracting significant interest due to its evident benefits including reduced latency, enhanced data privacy, and decreased connectivity reliance. While these advantages are compelling, they introduce unique challenges absent in traditional cloudbased approaches. In this paper, we delve into the intricacies of Edge-based learning, examining the interdependencies among: (i) constrained data storage on Edge devices, (ii) limited computational power for training, and (iii) the number of learning classes. Through experiments conducted using our MAGNETO system, that focused on learning human activities via data collected from mobile sensors, we highlight these challenges and offer valuable perspectives on Edge ML.

Air Quality Monitoring and Forecasting has been a popular research topic in recent years. Recently, data-driven approaches for air quality forecasting have garnered significant attention, owing to the availability of well-established data collection facilities in urban areas. Fixed infrastructures, typically deployed by national institutes or tech giants, often fall short in meeting the requirements of diverse personalized scenarios, e.g., forecasting in areas without any existing infrastructure. Consequently, smaller institutes or companies with limited budgets are compelled to seek tailored solutions by introducing more flexible infrastructures for data collection. In this paper, we propose an expandable graph attention network (EGAT) model, which digests data collected from existing and newly-added infrastructures, with different spatial structures. Additionally, our proposal can be embedded into any air quality forecasting models, to apply to the scenarios with evolving spatial structures. The proposal is validated over real air quality data from PurpleAir.

Air quality forecasting has garnered significant attention recently, with data-driven models taking center stage due to advancements in machine learning and deep learning models. However, researchers face challenges with complex data acquisition and the lack of open-sourced datasets, hindering efficient model validation. This paper introduces PurpleAirSF, a comprehensive and easily accessible dataset collected from the PurpleAir network. With its high temporal resolution, various air quality measures, and diverse geographical coverage, this dataset serves as a useful tool for researchers aiming to develop novel forecasting models, study air pollution patterns, and investigate their impacts on health and the environment. We present a detailed account of the data collection and processing methods employed to build PurpleAirSF. Furthermore, we conduct preliminary experiments using both classic and modern spatio-temporal forecasting models, thereby establishing a benchmark for future air quality forecasting tasks.

The United Arab Emirates (UAE) values its relationship with the desert, considering it a crucial part of its heritage and culture. However, the desert faces environmental challenges due to the improper disposal of garbage by visitors and the dumping of waste, as some perceive the desert as an empty wasteland. The rise in tourism exacerbates the problem, as litter negatively impacts the desert's ecology, wildlife, and natural habitats. Traditional litter collection methods involving human patrols are inadequate for the vast desert terrain. Drones equipped with high-resolution cameras offer a potential solution by conducting aerial surveys quickly and efficiently. However, the manual review of drone footage to detect litter is time-consuming. This paper explores the use of deep neural network architectures, such as Faster R-CNN, SSD, and YOLO, to develop litter detection models. These models focus on distinguishing litter from other man-made objects. The training dataset consists of thousands of samples, and the models are evaluated based on their performance in detecting and locating litter in drone images captured at different altitudes and environmental conditions. The evaluation includes objective and subjective analyses. The research aims to alleviate the practical challenges of litter detection in the desert by automating the process through computer vision-based object detection methods.

This paper proposes a novel mask-based output layer for multi-level hierarchical classification, addressing the limitations of existing methods which (i) often do not embed the taxonomy structure being used, (ii) use a complex backbone neural network with n disjoint output layers that do not constraint each other, (iii) may output predictions that are often inconsistent with the taxonomy in place, and (iv) have often a fixed value of n. Specifically, we propose a model agnostic output layer that embeds the taxonomy and that can be combined with any model. Our proposed output layer implements a top-down divide-and-conquer strategy through a masking mechanism to enforce that predictions comply with the embedded hierarchy structure. Focusing on image classification, we evaluate the performance of our proposed output layer on three different datasets, each with a three-level hierarchical structure. Experiments on these datasets show that our proposed mask-based output layer allows to improve several multi-level hierarchical classification models using various performance metrics.

A good understanding of the underlying mechanisms that govern identities on the Web is a key aspect for ensuring the privacy of users but also solving some ethical related problems. This paper proposes SpeculoLab, a platform implementing a strong and well defined experimental protocol for supporting research in the area of multiple identities in the (social) Web. The platform supports an end-to-end control of the experimentation process and, more importantly, allows personalizing and extending every part of the process. SpeculoLab is provided as an open source for the community for further improvements and reinforcement.

Data Integrity (DI) is the ability to ensure that a data retrieved from a database is the same as that stored and processed. It is a major component of data security and has an important role for the quality of decision making. While most of the existing approaches rely on a reinforcement of security mechanisms, e.g., access control or cryptography techniques, to ensure a high data integrity quality, we follow an end-user perspective to this end to complement the security approach. Simple, yet powerful, the proposed approach is promising in the inclusion of the end-user into the complex problem of DI.

It is largely known that objective criteria like profit and market-share drive the decisions of engineering business processes. However, there are cases where subjective criteria (e.g., reputation and attitude) need also to be taken into account, which will definitely impact the objective criteria. These cases fall into examining business processes from a social perspective. This paper discusses the mechanisms of making a business process's components (task, person, and machine) exhibit certain social qualities like selfishness and goodwill. This exposure is dependent on three criteria that are resource availability, transactional properties of tasks, and profit. An online system demonstrates the use of these criteria when capturing and quantifying the social qualities in business processes.

Database auditing is a prerequisite in the process of database forensics. Log files of different types and purposes are used in correlating evidence related to forensic investigation. In this paper, a new framework is proposed to explore and implement auditing features and DBMS-specific built-in utilities to aid in carrying out database forensics. The new framework is implemented in three phases, where ideal forensic auditing settings are suggested, techniques and approaches to conduct forensics are evaluated, and finally database forensic tools are investigated and evaluated. The research findings serve as guidelines toward focusing on database forensics. There is a crucial need to fill in the gap where forensic tools are few and not database specific.

In this paper, we introduce LAICOS, a social Web search engine as a contribution to the growing area of Social Information Retrieval (SIR). Social information and personalization are at the heart of LAICOS. On the one hand, the social context of documents is added as a layer to their textual content traditionally used for indexing to provide Personalized Social Document Representations. On the other hand, the social context of users is used for the query expansion process using the Personalized Social Query Expansion framework (PSQE) proposed in our earlier works. We describe the different components of the system while relying on social bookmarking systems as a source of social information for personalizing and enhancing the IR process. We show how the internal structure of indexes as well as the query expansion process operated using social information.

We present in this paper a contribution to IR modeling by proposing a new ranking function called SoPRa that considers the social dimension of the Web. This social dimension is any social information that surrounds documents along with the social context of users. Currently, our approach relies on folksonomies for extracting these social contexts, but it can be extended to use any social meta-data, e.g. comments, ratings, tweets, etc. The evaluation performed on our approach shows its benefits for personalized search.

In this paper, we present a contribution to IR modeling. We propose an approach that computes on the fly, a Personalized Social Document Representation (PSDR) of each document per user based on his social activities. The PSDRs are used to rank documents with respect to a query. This approach has been intensively evaluated on a large public dataset, showing significant benefits for personalized search.

This paper describes SONDY, a tool for analysis of trends and dynamics in online social network data. SONDY addresses two audiences: (i) end-users who want to explore social activity and (ii) researchers who want to experiment and compare mining techniques on social data. SONDY helps end-users like media analysts or journalists understand social network users interests and activity by providing emerging topics and events detection as well as network analysis functionalities. To this end, the application proposes visualizations such as interactive time-lines that summarize information and colored user graphs that reflect the structure of the network. SONDY also provides researchers an easy way to compare and evaluate recent techniques to mine social data, implement new algorithms and extend the application without being concerned with how to make it accessible. In the demo, participants will be invited to explore information from several datasets of various sizes and origins (such as a dataset consisting of 7,874,772 messages published by 1,697,759 Twitter users during a period of 7 days) and apply the different functionalities of the platform in real-time.

Extracting useful facts from large datasets has always been a challenging and critical issue for both research and industry. We present Vizpicious, a tool which borrows some ideas from social network analysis and semantic web to help investigators with such tasks, with a simple to use interface supporting them from the data integration phase until the analysis and the extraction of useful facts, and then provides more complex querying-based analysis capabilities.

In this paper, we describe our work in progress on Web services recommendation for services composition in a Mashup environment, by proposing a new approach to assist end-users based social interactions capture and analysis. This approach uses an implicit social graph inferred from the common composition interests of users. We describe the transformation of users-services interactions into a social graph and a possible means to leverage that graph to derive service recommendation. As this work is in progress, this proposal was implemented within a platform called SoCo where preliminary experiments show interesting results.

In this paper we address the problem of Web Mashups full completion which consists of predicting the most suitable set of (combined) services that successfully meet the goals of an end-user Mashup, given the current service (or composition of services) initially supplied. We model full completion as a frequent sequence mining problem and we show how existing algorithms can be applied in this context. To overcome some limitations of the frequent sequence mining algorithms, e.g., efficiency and recommendation granularity, we propose FESMA, a new and efficient algorithm for computing frequent sequences of services and recommending completions. FESMA also integrates a social dimension, extracted from the transformation of user-service interactions into user-user interactions, building an implicit graph that helps to better predict completions of services in a fashion tailored to individual users. Evaluations show that FESMA is more efficient outperforming the existing algorithms even with the consideration of the social dimension. Our proposal has been implemented in a prototype, SoCo, developed at Bell Labs.

We propose a new approach for social and personalized query expansion using social structures in the Web 2.0. While focusing on social tagging systems, the proposed approach considers (i) the semantic similarity between tags composing a query, (ii) a social proximity between the query and the user profile, and (iii) on the fly, a strategy for expanding user queries. The proposed approach has been evaluated using a large dataset crawled from del.icio.us.

With the emergence of Web 2.0 and the related technologies, composing services has left the traditional frontiers of enterprises. In fact, end-users need to use a certain kind of composition in different situations since Web 2.0 has brought a set of technologies making it easy to create or collaborate on new services or use others' services, e.g., mashups. On the other hand, users participate in different communities and social networks to share common interests and find expertise offered by others. Even if Web services composition tools like mashups include a community dimension helping the service creation process (tagging, rating, ...), they completely ignore this social dimension at the composition level. Our approach is to bring the generated knowledge from interactions between users and services (and by extension from social environments) to enhance services composition. This paper reviews some related concepts and work and then introduces a first view of our framework, named Social Composer (SoCo), aiming at handling this issue from a social networking perspective. SoCo provides dynamic recommendations for services discovery and selection based on the users' interactions and a social network implicitly built from the interactions between users and services, and the different services compositions operated in the user's social network as well as the global social network.

This paper presents a step in a long process of analyzing, structuring, and retrieving multimedia databases. Indeed, we propose to bring an improvement to an existing content based image retrieval approach. We propose an effective method for locally updating neighborhood graphs which constitute our multimedia index. This method is based on an intelligent way for locating points in a multidimensional space. Promising results are obtained after experimentations on various databases. Future issues of the proposed approach are very relevant