Recipient of the IEEE COMPSAC 2026 Best Paper Award.

In natural language processing, topic modeling aims to extract a corpora latent structure. In recent years, optimal transport distances have improved the topic extraction capabilities of Neural Topic Models (NTMs). More precisely, the Sinkhorn-Knopp algorithm is used to compute the blurred Wasserstein distance with relatively low complexity and is fully differentiable. This algorithm ease of implementation and advantages are thus particularly interesting for enforcing desired properties in NTMs. However, the algorithm can be unstable and inefficient under low blur setups, hence hindering overall topic model performances. In this article, we first assess the stability and efficiency of the Sinkhorn-Knopp algorithm in NTM scenarios. We compare five of the most relevant variations of this algorithm, and three distinct usages in NTMs. We evaluate each specific Sinkhorn-Knopp algorithm variation and topic model architecture independently, under various quantitative and qualitative metrics. Furthermore, we propose a novel method that focuses on the Sinkhorn-Knopp algorithm initialization, by reusing its dual variables from previous model updates as warm-start values. Our experiments reveal that our method can drastically improve the computation efficiency of the algorithm by reducing its number of iterations by up to 70%, and is easily applicable to any topic model using the Sinkhorn distance.

Federated Learning (FL) has emerged as a promising paradigm for decentralized machine learning, enabling multiple clients to collaboratively train a global model without sharing their raw data. Despite its privacy-preserving design, FL remains vulnerable to privacy leakage through inference attacks, such as membership inference, and to integrity threats like Byzantine behaviors that can degrade model reliability. To address these risks, we propose Local Data Privatization Preprocessing (LDPP), a lightweight client-side method that enforces differential privacy while preserving the statistical properties of local data. LDPP operates through a three-stage process: (i) transforming data into a standardized representation (normal or uniform), (ii) injecting calibrated noise using differential privacy mechanisms, such as Laplace or Gaussian distributions, and (iii) applying an inverse transformation to asymptotically recover the original data distribution. We formally prove that LDPP satisfies $\epsilon$-differential privacy and maintains key distributional characteristics. Additionally, LDPP can be combined with robust aggregation techniques, such as Krum, to strengthen defense against adversarial tampering. Comprehensive experiments in EMNIST and MedMNIST datasets demonstrate that LDPP significantly reduces the success of membership inference attacks and improves robustness under label-flipping scenarios while preserving high model accuracy. These findings position LDPP as a scalable and practical solution to improve both privacy and robustness in federated learning frameworks.

Issue Tracking Systems software serves as an interface between a company and its customers. Customers can report bugs and seek assistance, among other demands. Reported issues include textual description, along with company defined metadata, aim at simplifying issue treatment by experts. In the context of the rapid growth of customer-reported issues, the manual treatment process becomes tedious and time-consuming. As a result, more and more studies focus on automating parts of this process, using semantic extraction and topic modeling approaches to automatically classify issues. To this end, most approaches consider the issue of textual description along with metadata, which can be a source of uncertainty and misleading in many real-world scenarios. Besides, knowledge from the company experts is often neglected. In this paper, we propose a general taxonomy of information incorporation into topic models. This aims to assemble all existing techniques, to further detect literature gaps. In addition, we propose a technique to incorporate expert knowledge into neural topic models. We evaluate our techniques and others in the literature on a real-world dataset coming from the JIRA software of a French HR management company. Results show a significant increase of more than 22% in classification performances when using expert knowledge, in addition to the issue textual description. The results validate our approach's effectiveness in improving the automatic classification of issues.

Anomaly detection is a common task in various domains, which has attracted significant research efforts in recent years. Existing reviews mainly focus on structured data, such as numerical or categorical data. Several studies treated review of anomaly detection in general on heterogeneous data or concerning a specific domain. However, anomaly detection on unstructured textual data is less treated. In this work, we target textual anomaly detection. Thus, we propose a systematic review of anomaly detection solutions in the text. To do so, we analyze the included papers in our survey in terms of anomaly detection types, feature extraction methods, and machine learning methods. We also introduce a web scrapping to collect papers from digital libraries and propose a clustering method to classify selected papers automatically. Finally, we compare the proposed automatic clustering approach with manual classification, and we show the interest of our contribution.

Diabetes mellitus is a chronic disease that appears when the pancreas does not produce enough insulin or the body does not correctly use its insulin. If not adequately managed or diagnosed on time, this pathology can cause a lot of damage to the body organs, such as the heart, eyes, kidneys, and so on. Research carried out through machine learning has made it possible to have increasingly efficient and precise models for detecting and preventing type 2 diabetes. However, most of the models mentioned do not offer guarantees on the privacy of patient data used during the training process. In addition, these models are generally stored in a centralized repository, where the analysis is performed with full access to sensitive content, implying increased attack risks on confidentiality and privacy. This paper proposes a Differentially Private Stochastic Gradient Descent applied to the Federated Averaging (DPSGDFedAvg) model for diabetes prediction using the Pima Indian dataset. In first results, we obtained an accuracy between 60% and 70% with a raised level of privacy. We demonstrate in this work the feasibility and effectiveness of the DPSGDFedAvg model in offering a raised level of privacy and maintaining utility of the global FL model.

The world has been experiencing several crises recently, particularly on the social front. Therefore, new technologies have been adapted to provide the most diverse possible solutions, including crowdfunding platforms that concentrate on social projects. They have recently piqued the interest of investors and donors, particularly those based on blockchain technology, thanks to their ability to achieve reliability. Social crowdfunding platforms have developed new strategies for luring donations and investments. However, there is still a lack of development of these ideas and exploiting the benefits and services provided by blockchain technology properly. This paper presents blockchain technology in a socially oriented crowdfunding platform reward-based that aims to provide a transparent, secure, auditable, and efficient system. BELONG is the first leading platform that merged the ideas of crowdfunding, donations, and charitable investments with a type of blockchain-based token called Non-fungible tokens (NFTs). The goal is to create safe investment channels, and that is because of the dearth of studies on the idea of integrating NFTs into humanitarian, charitable, or social activities. It relies on two strategies for seeking funds; the bedrock on which all two are built is the NFT. This study intends to reach out to all societal stakeholders interested in this field. As a result, each strategy targets a specific category, including donors, investors, and individuals, to make funding opportunities available for everyone. A dedicated prototype, using Ethereum and Vuejs, is implemented to demonstrate the platform's feasibility.

The Internet of Things (IoT) is expected to improve the individuals’ quality of life. However, ensuring security and privacy in the IoT context is a non-trivial task due to the low capability of these connected devices. Generally, the IoT device management is based on a centralized entity that validates communication and connection rights. Therefore, this centralized entity can be considered as a single point of failure. Yet, in the case of distributed approaches, it is difficult to delegate the right validation to IoT devices themselves in untrustworthy IoT environments. Fortunately, the blockchain may provide decentralization of overcoming the trust problem while designing a privacy-preserving system. To this end, we propose a novel privacy-preserving IoT device management framework based on the blockchain technology. In the proposed system, the IoT devices are controlled by several smart contracts that validate the connection rights according to the privacy permission settings predefined by the data owners and the stored record array of detected misbehavior of each IoT device. In fact, smart contracts can immediately detect the devices that have vulnerabilities and have been hacked or pose a threat to the IoT network. Therefore, the data owner’s privacy is preserved by enforcing the control over the own devices. For validation purposes, we deploy the proposed solution on a private Ethereum blockchain and give the performance evaluation.

The Internet of Things (IoT) provides the opportunity to collect, process and analyze data. This opportunity helps to understand preferences and life patterns of individuals in order to offer them customized services. However, privacy has become a significant issue due to the personal nature of the knowledge derived from these data and the involved potential risks. Despite the increasing legislation pressure, few proposed solutions have dealt with the privacy requirements, such as consent and choice, purpose specification, and collection limitation. In this paper, we propose a privacy ontology in order to incorporate privacy legislation into privacy policies while considering several privacy requirements. Our proposed ontology aims both at making the smart devices more autonomous and able to infer data access rights and enforcing the privacy policy compliance at the execution level. We implemented and evaluated our privacy ontology based on a healthcare scenario.

Nowadays, cloud computing is an emerging technology thanks to its ability to provide on-demand computing services (hardware and software) with less description standardization effort. Multiple issues and challenges in discovering cloud services appear due to the lack of the cloud service description standardization. In fact, the existing cloud providers describe, their similar offered services in different ways. Thus, various existing works aim at standardizing the representation of cloud computing services while proposing ontologies. However, since the existing proposals were not evaluated, they might be less adopted and considered. Indeed, the ontology evaluation has a direct impact on its understandability and reusability. In this paper, we propose an evaluation approach to validate our proposed Cloud Service Ontology (CSO), to guarantee an adequate cloud service discovery. This paper contribution is threefold. First, it specifies a set of patterns and anti-patterns in order to evaluate CSO. Second, it defines an anti-pattern detection method based on SPARQL queries which provides a set of correction recommendations to help ontologists revise the ontology. Finally, some experiment tests were conducted in relation to: (i) the method efficiency and (ii) anti-pattern detection of design anomalies as well as taxonomic and domain errors within CSO.