ABSTRACT. Alternative Polyadenylation is an essential post-transcriptional process. By selecting which polyadenylation (poly(A)) site is cleaved, a single gene is enabled to generate multiple transcript isoforms with different 3′ untranslated regions. However, previous studies have focused on the identification of poly(A) sites within genomic sequences, with limited research dedicated to quantifying the usage level of alternative poly(A) sites in the same gene. In this paper, we propose a novel Iterative Refinement model for predicting the usage level of each poly(A) site within a gene, namely RepAs. In particular, the model iteratively predicts a usage value for each poly(A) site, while simultaneously refining the predictions for the previous sites within the same gene, to account for the competing interactions between them. Furthermore, RepAs integrates a feature refinement process with a custom attention mechanism to progressively improve the representation of poly(A) site-related features. Our experiments on publicly available datasets demonstrate the effectiveness of the proposed approach, outperforming other state-of-the art methods. Finally, for reproduction purposes, we make the implementation of RepAs publicly available at https://anonymous.4open.science/r/RepAs-5BC8.

ABSTRACT. Biclustering is a powerful tool for analyzing gene expression time series, providing the ability to simultaneously explore both gene and condition dimensions. Unlike traditional clustering methods, which are limited to one dimension, biclustering uncovers local patterns of co--expression, making it particularly well--suited for the analysis of dynamic gene expression data. In this paper, we introduce BicTree, a novel tree--based algorithm designed to efficiently identify temporally coherent biclusters in gene expression time--series data. BicTree employs a three--step process: Plaid Filtering for feature elimination, Tree Extension to partition gene sets based on correlation, and Bicluster Extraction to identify significant patterns.

We evaluate BicTree performance on the yeast cell cycle dataset and compare it with two established biclustering algorithms, e--CCC and CC--TSB. Our results demonstrate that BicTree achieves a superior balance between computational efficiency and bicluster quality, producing highly coherent biclusters with an average Pearson correlation coefficient of 0.91. In contrast, e--CCC and CC--TSB display lower correlations and higher instability, particularly in the presence of noise. Furthermore, BicTree performs well in detecting biologically relevant biclusters, as validated by gene ontology enrichment analysis. These results highlight the effectiveness of BicTree in discovering biologically meaningful gene expression patterns, offering a robust and efficient solution for the analysis of time--series data in genomics.

ABSTRACT. Predicting whether a peptide will be presented on Major Histocompatibility Complex class I (MHC-I) molecules has significant implications for vaccine design. Existing work provides numerous deep learning-based predictors for peptide presentation on MHC-I molecules with high accuracy. However, these predictors are predominantly black-box functions, offering minimal insight into their decision-making processes. To address this issue and concerns raised in AI regulations like EU AI Act, we explore a transparent and interpretable model for predicting peptide MHC-I binding. Our model, MHCRule uses a rule-based approach that learns rules for binding prediction based on the input sequences and the biochemical properties of the amino acids in the peptide sequence.

When comparing to state-of-the art predictors, we find MHCRule gives comparable performance but with the added benefit of explaining each prediction output with interpretable rules. Additionally, we find quality of the rules from MHCRule aligns with biological expectations and known binding motifs for MHC-I alleles.

ABSTRACT. Deep learning-based medical image segmentation has rapidly gained attention in analyzing complex medical images, promising to enhance diagnostic precision and treatment planning. However, training deep learning models on high-dimensional data presents substantial challenges, particularly regarding computational demands related to memory usage and processing time. Data downsampling is a widely employed strategy that reduces memory requirements and accelerates training processes to mitigate these issues. Conventionally, nearest neighbor (NN) interpolation has been utilized to downsample ground truth labels. However, this approach often leads to loss of class information and can detrimentally impact segmentation performance compared to training on the original high-dimensional data. This study introduces a novel downsampling method, Probabilistic K-Nearest Neighbors (PKNN), specifically designed for medical image segmentation. The PKNN method effectively minimizes pixel loss while preserving critical segmentation information. We conducted extensive experiments using various values of K at both half and quarter resolutions. The results demonstrate that PKNN significantly outperforms NN interpolation, achieving improvements in the Intersection over Union (IoU) metrics of 1.33% and 2.29% on the Abdominal CT dataset and 2.88% and 2.44% on the CVC-ClinicDB dataset. Notably, the performance of PKNN approaches that of models trained on original high-resolution data, underscoring its effectiveness in maintaining segmentation accuracy despite the reduced resolution.

ABSTRACT. The increasing application of information and communication technologies across various fields has led to the emergence of numerous heterogeneous systems, creating complex digital ecosystems where data integration and interoperability present significant challenges. This paper proposes a comprehensive framework for automated user attribute mapping that addresses these challenges, leveraging bidirectional encoder representations from transformers (BERT), a state-of-the-art natural language processing (NLP) model, to enable accurate mapping across systems with diverse data representations. The framework integrates a metadata registry (MDR) and blockchain to securely manage and share mapped information, ensuring consistent data integration across distributed environments. By overcoming the limitations of traditional rule-based methods, the proposed framework significantly improves system interoperability. The experimental results are provided to demonstrate the feasibility of the proposed mapping method, showcasing its practical applicability in real-world scenarios. These results confirm the framework’s potential to enhance data integration and interoperability in complex, heterogeneous systems.

ABSTRACT. Controlled Environment Agriculture (CEA) is increasing attention as a solution to food security challenges posed by population growth and climate change. Current CEA systems primarily manage plant growth through sensor-based environmental data. Moreover, previous research for the CEA system predominantly focuses on improving system-oriented interoperability. This contributes to the efficiency of managing CEA systems, but there are difficulties in directly analyzing the growth status of plants. To address this, we propose a system that can directly monitor plants' status by utilizing their electrical signals. This system aims to improve the interoperability of growing environments between CEA systems by classifying environmental conditions based on electrical signals. Experimental results showed an accuracy of 0.9802 in evaluating environmental consistency between CEA systems. Lastly, we validate the method for improving the interoperability of CEA systems using a status diagram and performance verification based on scenarios.

ABSTRACT. The use of knowledge graphs in recommender systems has become one of the common approach to addressing data sparsity and cold start problems. Recent advances in large language models (LLMs) offer new possibilities for processing side and context information within knowledge graphs. However, consistent integration across various systems remains challenging due to the need for domain expert intervention and differences in system characteristics. To address these issues, we propose a consistent approach that extracts both general and specific topics from side and context information using LLMs. First, general topics are iteratively extracted and updated for both side and context information. Then, specific topics are extracted and replaced using context information. Finally, a refining algorithm handles synonymous topics that may arise during specific topic extraction. This approach allows general topics to capture broad knowledge across a wide range of item characteristics, while specific topics focus on detailed attributes, providing a more comprehensive understanding of items' semantic features. Furthermore, our approach is constructed using standardized metagraph, eliminating the need for domain expert intervention and ensuring interoperability. Experimental results demonstrate that our method significantly improves recommendation performance across different knowledge graphs in various domains.

ABSTRACT. Digital twins (DTs) serve as precise virtual representations of real-world objects, processes, or systems, enabling simulation, monitoring, and advanced analysis to enhance operational insights and support decision-making. A relevant challenge to the progress of DTs is achieving interoperability among their components. This limitation hinders the seamless integration of DTs with other technologies and platforms, reducing their applicability. In this context, this paper presents the results of a systematic mapping study offering a comprehensive overview of existing approaches for interoperability in DTs. It explores their goals, the achieved levels of interoperability, and application domains, providing insights for researchers and practitioners in the field.

ABSTRACT. With the growing reliance on social media platforms as primary sources of information, users increasingly face the challenge of efficiently managing and retrieving relevant information. Given the large volume of messages users receive daily across multiple groups, current research is being conducted to develop systems that can effectively detect emerging conversations about topics of interest. This work proposes a novel approach to enhancing the user experience in Telegram by developing a topic-based identification of new conversations.

A two-step approach is employed. First, a conversation model identifies whether a message begins a new conversation. Then, a topic model assigns topics to the conversations based on user preferences. The system has been validated using real-world data collected from university-related Telegram groups, including six recurring topics such as exams, homework, and deadlines. The proposed models perform well in identifying both new conversations and relevant topics, with F1 score reaching 0.94% and macro F1 score of 0.68%, respectively. The proposed method provides a lightweight, general-purpose solution that can be easily adapted to other messaging platforms, ultimately improving message management and user engagement.

ABSTRACT. Hierarchical topic modeling is a well-known technique for deriving comprehensive insights on a given dataset. However, it is challenging to choose the best-suited hierarchical topic model among many candidates, given that the model generally depends on the dataset under analysis. Moreover, even that chosen model typically produces an overwhelming number of leaf topics, making it hard to correctly interpret the result derived from the model. Although topic coherence is an often used metric to assess the quality of a model, coherence cannot reflect the unique characteristics of the hierarchical structure when applied to the model as well. To address these concerns, we propose a novel evaluation metric, HARIN (HierArchical haRmony INdex). The proposed HARIN metric effectively reflects the overall topic coherence, diversity, and similarity among parent-child and sibling layers in the produced hierarchy. We test the validity of HARIN by comparing it against competing metrics and human assessments of the quality of topic hierarchies from four popular models applied to five real-world datasets. Our experiments demonstrate that HARIN achieves approximately 85% accuracy in model ranking compared to human scores, surpassing the leading competing metric, coherence, by 1.4×. Notably, HARIN's mean reciprocal rank of 1 highlights its exceptional ability to recommend the optimal hierarchical topic model. We show that HARIN outperforms the coherence metric in identifying the best model while also pruning the right number of leaf topics, thereby enhancing both model selection and result interpretability.

ABSTRACT. Personalized news recommendation has become an essential technology for online news services. For effective personalized news recommendation, it is ideal to utilize various information, such as the title and body text. However, some news services do not retain rich information such as the body text, and only titles may be available. Large Language Models (LLMs) such as ChatGPT have attracted much attention in the field of Natural Language Processing (NLP) owing to their excellent sentence understanding and generation capabilities. In this paper, we propose a news recommendation framework that utilizes data augmentation with ChatGPT. By inputting our prompt and news title into ChatGPT, we extend the information in news articles to supplement the news content feature. In particular, we focus on two directions of title extension: (1) user direction and (2) content direction. In the proposed framework, ChatGPT infers and outputs (1) the target audience of the news article in the user direction, and (2) the categories of the news article in the content direction. These output sentences are then used as input to the news recommendation model, along with the title of the news article. Evaluation experiments on a real-world news service dataset show that the proposed framework outperforms conventional methods by up to 1.65\% in AUC (area under the ROC curve). These findings highlight the importance of extending the content features of news articles from their titles and utilizing them for recommendation through various prompting strategies. We have provided our code and GPT-generated data to enable other researchers to replicate our findings.

ABSTRACT. Social networks have become invaluable data sources, enabling many studies about human behavior, language, and social prediction. Users commonly propose new expressions to make communication faster, more fluid, and more original. Moreover, although they follow a specific idiom, the employed language is also mixed-coded, making even the same idiom go beyond the usual variations. This way, texts in Portuguese from different regions that already have linguistic, syntactic, lexical, and orthographic differences go to a new level when written in social media posts. This paper proposes relying on Portuguese language models to analyze the variations of Portuguese in social media posts in Brazil, Portugal, and Mozambique. Our methodology involves pre-trained BERT-based models tuned for the task of text classification and the analysis of the attribution value of the models in the tokens. Moreover, we propose inspecting the ability of the models to predict masked terms specific to regional variants. The analysis reveals that models trained on one variation of Portuguese can perform well on more than one variation. However, there are still challenges in distinguishing between the Portuguese and Mozambican variants.

ABSTRACT. Recent media have increasingly shifted towards multimedia formats that simultaneously utilize visual and linguistic information. Re-search on multimodal AI is actively conducted to analyze large-scale multimodal data effectively. Multimodal AI fuses the probability and feature values extracted from single modalities by a backbone model, enabling the simultaneous analysis of multimodal information. This allows for discovering new insights that may not be detectable through single-modality analysis. Depending on the data collection environment, multimedia can be classified into one-to-one and one-to-many modality balances. Previous multimodal AI approaches analyze these one-to-many relationships by downsampling or duplicating data to fit a one-to-one relationship. In this paper, we optimize multimedia analysis in one-to-one and one-to-many modality balances based on the local and global context analysis capabilities of multimodal AI and the multimodal analysis characteristics of backbone models. The multimedia analysis system employs late score and feature fusion to independently analyze the local context as the baseline for multimodal AI. In contrast, early and hierarchical feature fusion is utilized for comprehensive global context analysis. The backbone models used include ViT and RoBERTa to analyze the overall structure of multimodal data and BEiT and DeBERTa to analyze structural features. Experimental results show that, in the duplication method, late score and feature fusion, which independently analyze the local context of multimodal data, are 0.56% more accurate and achieve an f1 score that is 0.025 higher. Additionally, BEiT and DeBERTa, which analyze structural features, demonstrate a 0.2% increase in accuracy and a 0.0167 improvement in f1 score. In the downsampling method, early and hierarchical feature fusion, which comprehensively analyzes the global context, outperforms by 1.17% in accuracy and 0.0164 in f1 score. Furthermore, ViT and RoBERTa, which focus on analyzing the overall structure, exhibit a 0.74% improvement in accuracy and a 0.0064 increase in f1 score.

ABSTRACT. Bigraphs are a formal model for representing (ubiquitous) systems with strong notations of both space, e.g. a person in a room, and non-spatial relations, e.g. mobile phone communication regardless of location. They have been used in a wide range of scenarios including sensor systems, IoT configuration languages, and communications protocol design. While implementations of the bigraph theory exist, e.g. BigraphER, until now, there has been no attempt to formalise the theory in a theorem prover. We show an implementation of the bigraph theory in the Coq theorem prover, including the main bigraph type specification and common manipulation operators, e.g. composition and tensor product. This is a key step to fully formalising the theory and paves the way for a certified implementation for use in safety critical scenarios.

ABSTRACT. Game semantics provides fully abstract (sound and complete) models for open program fragments with undefined, non-local, identifiers (e.g. library functions). This is achieved by using the ``most general'' models for undefined identifiers, i.e. the most generic context in which the program fragment will be inserted. Given a safety property as an assertion, we want to find the most permissive models of undefined identifiers, i.e. the weakest safe context, that are sufficient to ensure safety of the given program fragment.

To solve this problem, we present a novel approach that uses symbolic game semantic models of open program fragments and logical abduction. By using symbolic values instead of concrete ones for integers, we represent algorithmic game semantics of program fragments with unbounded integers as finite-state symbolic automata (regular languages) that are amenable for automatic reasoning. Then, we reduce the problem of inferring minimal assumptions on the behaviors of undefined identifiers in order to prove safety of the given program fragment to a logical abduction task. We evaluate our approach on a range of examples and demonstrate its ability to synthesize weakest safe contexts.

ABSTRACT. The analysis of timed properties of programs is a complex task, as it is highly dependent on both the software and the hardware. In this work, we propose a framework for modeling with timed formal models the execution of programs, taking into account the micro-architecture of the machine on which it executes. We model both the program, at the instruction set architecture level, and the hardware, including the processor micro-architecture, using time Petri nets. Our implementation uses the ARM Cortex-M instruction set architecture and a hardware architecture representative of microcontrollers used in IoT nodes. The whole translation is fully automated and allows, starting from binary code, to automatically produce the models usable by the state-of-the-art time Petri net model checker Roméo. In addition, and as a proof of concept, we show how we can check, and enforce to some extent, opacity properties on programs, leveraging the ability of Roméo to verify a parametric timed extension of the classical computation tree logic.

ABSTRACT. Embedded automotive software development increasingly uses formal methods to guarantee correctness and dependability. Generally, we can use models of high-level (hardware-independent) specifications to generate high-level code. However, we need to model low-level (hardware-dependent) specifications individually for each hardware architecture, which significantly increases modeling complexity and verification effort and limits the re-usability of these low-level models to generate hardware-specific code. The problem is even more significant when dealing with complex low-level functionalities (e.g., context switch, system initialization) of an AUTOSAR OS.

In this paper, we propose a refinement-based formal modeling approach to address this problem. First, we use UML-B and Event-B formal methods to implement a hardware-generic model of the low-level specification, which abstracts hardware-specific details. Next, we refine the generic model by including the hardware-specific details in the final refinements. We verify the hardware-generic and -specific models using theorem proving and Linear Temporal Logic (LTL) model checking. This way, our approach facilitates portability and re-usability while ensuring the correctness of low-level functionalities. As proof of concept and evaluation, we model and verify the context switch of an AUTOSAR OS and refine the hardware-generic model of the use case for AURIX TriCore and RISC-V architectures.

ABSTRACT. As Large Language Models (LLMs) become increasingly integrated into digital communication, their impact on the sentiment of text has become a critical area of investigation. This study explores the effects of iterative rephrasing by LLMs, particularly focusing on how repeated processing with neutral prompts influences sentiment expression. Our research identifies distinct patterns of sentiment drift, revealing that open-source LLMs tend to shift towards a positive sentiment over multiple iterations, which may lead to a homogenization of emotional expression in digital discourse. In contrast, advanced models such as GPT-4o demonstrate a superior ability to preserve the original sentiment across iterations. To quantify these effects, we introduce the Sentiment Fidelity Score, a novel metric that assesses the capacity of LLMs to maintain emotional tone through successive rephrasings. These findings offer valuable insights into the design and deployment of LLMs in applications where the preservation of sentiment is crucial, highlighting both the strengths and limitations of current models.

ABSTRACT. The use of online services for advertising job positions has grown in the last decade, thanks to the ability of Online Job Advertisements (OJAs) to observe the labour market in near real-time, predict new occupation trends, identify relevant skills, and support policy and decision-making activities. Unsurprisingly, 2023 was declared the Year of Skills by the EU, as skill mismatch is a key challenge for European economies. In such a scenario, machine learning-based approaches have played a key role in classifying job ads and extracting skills according to well-established taxonomies. However, the effectiveness of ML depends on access to annotated job advertisement datasets, which are often limited and require time-consuming manual annotation. The lack of OJA annotated benchmarks representative of the real online OJA and skills distributions is currently limiting advances in skill intelligence.

To deal with this, we propose JobGen, which leverages large language models (LLMs) to generate synthetic OJAs. We use real OJAs collected from an EU project and the ESCO taxonomy to represent job market distributions accurately. JobGen enhances data diversity and semantic alignment, addressing common issues in synthetic data generation. The resulting dataset, JobSet, provides a valuable resource for tasks like skill extraction and job matching and is openly available to the community.

ABSTRACT. Generative large language models (LLMs) have revolutionized natural language processing (NLP) by demonstrating exceptional performance in interpreting and generating human language. There has been some exploration of their application to non-linguistic tasks, which could lead to significant advancements in fields that rely heavily on structured data and specialized knowledge. However, there has been limited direct comparison of the effects of model adaptation techniques for non-linguistic compared to linguistic tasks with LLMs. To this end, the work in this paper investigates the effects of fine-tuning and few-shot learning on pre-trained LLMs for non-linguistic tasks using chess puzzles as a case study task. We compare the impact of fine-tuning and few-shot learning on models performing the same task represented in both chess notation (i.e., non-linguistic data) and natural language descriptions of the same chess notations (i.e., natural language data). Our experiments with Mixtral-8x7B-v0.1 and Meta-Llama-3-70B resulted in a 5% lower average increase in performance after fine-tuning for non-linguistic tasks compared to linguistic tasks. Similarly, few-shot learning on pre-trained models exhibited a 3% lower average increase in performance for on non-linguistic tasks compared to linguistic tasks. Furthermore, few-shot learning on fine-tuned models resulted in a significant accuracy drop, particularly for Mixtral, with a 24.82% decrease for non-linguistic tasks. These results suggest that fine-tuning and few-shot learning for generative LLMs have stronger effects on linguistic tasks and their data than for non-linguistic.

ABSTRACT. Mixture of Experts (MoE) architectures have been increasingly employed to enhance the inference speed of Large Language Models (LLMs). However, in previous MoE approaches, there is no clear separation between the specialties of each expert, as they divide tasks based on patterns in the data rather than on specific topics. This limits their ability to specialize in particular languages or domains. Incorporating modularity by developing domain-specific experts within an MoE framework could enable more flexible and efficient LLMs, where modules can be selected and combined based on specific application needs, similar to software libraries. We present a Modular Mixture of Experts (MMoE) architecture that enhances efficiency and specialization in language models by integrating Knowledge Distillation (KD) with MoE to create domain-specific experts. In our approach, LLMs are compressed into smaller, domain-specific experts using KD, which are then combined within a modular MoE framework. In our experiments, the domains considered were English, German, French, and Python. This modularity allows for the flexible selection and deployment of any subset of experts tailored to specific application needs, facilitating resource-efficient deployment. A router directs inputs to the appropriate experts, ensuring that each input is processed by the most relevant specialized module. We evaluated different MoE configurations and found that our modular approach effectively handles multi-domain inputs, mitigates catastrophic forgetting, and preserves knowledge across multiple domains. We have open-sourced our dataset and codebase to facilitate further research in this domain.

ABSTRACT. Automatic video colorization is a challenging task where multiple plausible colorizations can be deployed for any black-and-white film. For single photos, it is possible to have human knowledge guide the colorization process through text prompts, but for all of the frames and entities shown in a video, it becomes more difficult to achieve. With recent advances in automatic video colorization, natural language processing and knowledge enrichment, it is feasible to leverage external knowledge in automatic text-guided video colorization. To realize this possibility, we propose RAGCol, a knowledge-enriched video colorization system which adapts the retrieval augmented generation (RAG) framework to an automated colorization pipeline. We validated our RAGCol on the DAVIS and Videvo datasets. RAGCol demonstrated an average improvement of 9% over the previous state-of-the-art L-CAD across the PSNR, SSIM, FID and FVD metrics. In a user study, we found that videos colorized by RAGCol were preferred by 74% on average over contemporary colorizers by human evaluators.

ABSTRACT. The social impact and global health crisis caused by the coronavirus (COVID-19) since late 2019 led to the development of a novel bio-inspired algorithm. This algorithm simulates the behavior and spread of the virus, known as the Coronavirus Optimization Algorithm (CVOA). It provides several advantages over similar approaches and serves as a basis for generalizing pattern or association identification from numerical datasets. In this study, essential updates and modifications are proposed to adapt the CVOA algorithm for mining numerical association rules. These changes involve adjustments to the encoding of individuals and the infection/mutation process. Additionally, parameter values are updated, and a new fitness function is proposed to be maximized. The main objective is to obtain high-quality numerical association rules for any dataset regardless of the number and range of attributes in the dataset. The implemented algorithm is compared to others designed for mining quantitative association rules in order to validate the results. For this reason, different datasets from the BUFA repository are used, confirming that CVOA is a promising option for discovering interesting association rules within numerical datasets.

ABSTRACT. In this work, we present a privacy-preserving framework for credit scoring systems deployed on Machine Learning as a Service (MLaaS) platforms. Our approach integrates an obfuscator-classifier model that enhances privacy while maintaining high accuracy for loan default prediction tasks. The obfuscator transforms sensitive financial data into a privacy-protected representation, minimizing the risk of privacy leakage and input reconstruction during inference. By employing a combination of center loss and noise addition, our model ensures a robust balance between privacy and utility.

Through extensive experiments, we demonstrate the effectiveness of our solution in reducing information leakage. For instance, our method achieves a 95.05% reduction in the average R2 score of reconstruction attacks, from 0.921 to 0.045. At the same time, we maintain high prediction accuracy, with only a negligible loss of 1.06% in public task accuracy, despite the added noise. These results highlight the scalability and adaptability of our framework for financial MLaaS applications, providing strong privacy protection without significantly compromising model performance.

ABSTRACT. This paper proposes a two-step approach to efficiently calibrate the stochastic models for the foreign exchange (FX) barrier options by utilising the predictive and generative power of machine learning. To tackle the limited availability of market prices from brokers, we propose a framework to first augment the model parameters via a generative model, the Variational Autoencoder Generative Adversarial Network (VAE-GAN) model, and then calibrate the model parameters with neural networks to approximate the mapping between synthetic market data and model parameters. In this work, we examine the performance of our two-step calibration approach by comparing it with the traditional calibration process in terms of robustness and efficiency. We evaluate our calibration using two performance metrics employed by major financial institutions. The results indicate that our method not only speeds up the calibration process from hours to seconds but also increases the variety of the dataset, covering a broader range of market conditions. Finally, we use the values output by our calibration method as initial values in the traditional calibrator. This approach helps the traditional calibrator achieve optimal parameters by either improving running time or the quality of the solutions, resulting in a closer match between the model-implied prices and the market prices.

ABSTRACT. Enriching sentences with qualitative knowledge is crucial for enhancing sentiment prediction and making the most of the available labelled data for training models. This is particularly important in domains like the financial one, where texts are usually brief and contain much-implied information. In this article, we introduce FLEX (Financial Language Enhancement with Guided LLM Execution), an automated system capable of retrieving information from a Large Language Model (LLM) to enrich financial sentences, making them more knowledge-dense and explicit. FLEX generates multiple potentially enhanced sentences and uses a new logic to determine the most suitable one. Since LLMs may introduce hallucinated answers, we have significantly reduced this risk by developing a new algorithm that selects the most appropriate sentences. This approach ensures that the meaning of the original sentence is preserved, avoids excessive syntactic similarity between versions, and achieves the lowest possible perplexity. These enhanced sentences are more interpretable and directly useful for downstream tasks like financial sentiment analysis (FSA). Compared to state-of-the-art methods, FLEX shows improvements in the accuracy of processing FSA tasks.

ABSTRACT. In industrial settings, the Software Team Formation Problem (STFP) is often hindered by inconsistency, bias, and overlooked collaboration history in manual team formation processes. These limitations create inefficiencies in aligning technical and collaborative skills. To address this, Genetic Algorithms (GAs) have been introduced, but traditional crossover methods restrict diversity and can lead to premature convergence, forming suboptimal teams. This paper introduces a GA designed to integrate both technical expertise and collaborative performance. By utilizing a graph-based fitness function and the novel Convex Combination Crossover technique, the algorithm explores team configurations more effectively. We validated our approach using data from 47 software projects and 149 developers across 12 simulation scenarios. The results show that the collaboration graph reliably captures team dynamics, while the Convex Combination Crossover outperforms traditional methods (Partially Mapped Crossover and One-Point Crossover) in generating more diverse and fit team structures. Our GA consistently improved team diversity, fitness, and overall configuration quality. These findings suggest that combining technical and collaborative factors leads to better decision-making in team formation, providing a more effective and scalable solution for industrial software projects

ABSTRACT. Hyperledger Fabric is a well-known framework for developing enterprise blockchain solutions. Developers of these blockchains must ensure the correct execution of read and write operations so that the smart contracts' application logic is consistent with the business logic. In this paper, we present a static analysis approach based on abstract interpretation to detect read-write set issues in Hyperledger Fabric smart contracts and avoid bugs and critical errors that could compromise blockchain applications. The analysis is implemented in GoLiSA, a semantics-based static analyzer for Go applications. Our experimental results show that the proposed analysis can detect read-write set issues on a significant benchmark of existing applications. Moreover, it achieves better results when detecting read-after-write issues than other well-known open-source analyzers for Hyperledger Fabric smart contracts.

ABSTRACT. A blockchain is a trustless system in an environment populated by untrusted peers. Code deployed in blockchain as a smart contract should be cautious when invoking contracts of other peers as they might introduce several risks and unexpected issues. This paper presents an information flow-based approach for detecting cross-contract invocations to untrusted contracts, written in general-purpose languages, that could lead to arbitrary code executions and store any results coming from them. The analysis is implemented in GoLiSA, a static analyzer for Go. Our experimental results show that GoLiSA is able to detect all vulnerabilities related to untrusted cross-contract invocations on a significant benchmark suite of smart contracts written in Go for Hyperledger Fabric, an enterprise framework for blockchain solutions.

ABSTRACT. Deploying smart contracts and invoking their functions on blockchains incur gas costs, which depend on the operations executed by those functions. This makes optimizing the gas cost of smart contract functions a rewarding goal. However, existing approaches to gas cost optimization of smart contracts mainly involve rule-based optimization or automatic optimization for specific types of patterns. In this paper, we discuss a novel approach to automatically retrieving optimized versions of Solidity functions from a repository of smart contracts. The system identifies and suggests gas-efficient alternatives that maintain functional equivalence by comparing the opcode sequences of individual functions. We evaluate this approach on a dataset of 16,529 functions from real-world contracts, demonstrating substantial gas savings, as high as 34% on average when considering the most similar functions.

ABSTRACT. To enable market participants to trade tokens in cryptoeconomic systems, automated market makers (AMMs) typically rely on liquidity providers (LP) that deposit tokens in exchange for rewards. To profit from such rewards, LPs must use effective liquidity provisioning strategies. However, LPs lack guidance for developing such strategies, which often leads them to experience financial losses. We developed a measurement model based on impermament loss to analyze influences of key parameters (i.e., liquidity pool type, position duration, position range size, and position size) of liquidity provisioning strategies on LPs' returns. To reveal influences of those key parameters on LPs' profits, we applied the measurement model to analyze 700 days of historical liquidity provision data of Uniswap v3. By uncovering the influences of key parameters that constitute different liquidity provisioning strategies on profitability, this work supports LPs in developing provisioning strategies that yield higher returns.

ABSTRACT. Technology parameter maps summarize experiences with specific parameters in production processes, e.g., milling, and significantly help in designing new or improving existing production processes. Businesses could greatly benefit from globally exchanging such existing knowledge across organizations to optimize their processes. Unfortunately, confidentiality concerns and the lack of appropriate designs in existing data space frameworks—both in academia and industry—greatly impair respective actions in practice. To address this research gap, we propose MapXchange, our homomorphic encryption-based approach to combine technology parameters from different organizations into technology parameter maps while accounting for the confidentiality needs of involved businesses. Central to our design is that it allows for local modifications (updates) of these maps directly at the exchange platform. Moreover, data consumers can query them, without involving data providers, to eventually improve their setups. By evaluating a real-world use case in the domain of milling, we further underline MapXchange's performance, security, and utility for businesses.

ABSTRACT. Federated data spaces allow organizations to share and control their own data across various domains, but their exposure to cyber attacks has increased due to a surge in newly discovered vulnerabilities. Existing solutions to secure them focus on messaging protocol protection (e.g., using cryptographic means), but this is not sufficient. Attackers may exploit additional vulnerabilities to cause significant issues (e.g., disrupting the availability of services). To this end, we propose SHIELD, a security-by-design approach for federated data spaces, which leverages attack graphs and trust computation to mitigate the risks of cyber attacks. Mitigation is accomplished by proactively assessing the data spaces' weaknesses and implementing security messaging measures to prevent detrimental attacks. A prototype implementation of SHIELD using publish-subscribe as a messaging mechanism is experimentally evaluated over a real architecture in a V2X (Vehicle-to-Everything) scenario.

ABSTRACT. Inspired by the Sustainable Development Goals of the United Nations, the European Union (EU) has taken numerous regulatory measures to realize a low-emission circular economy (CE). In particular, the EU is advancing the idea of the Digital Product Passport (DPP) to promote data sharing between organizations and improve the environmental impact of products on the basis of the data. Although the exact requirements for the content of the DPP are yet to be defined, the global climate agenda suggests that across all product groups the carbon footprint will be an important part of the DPP. Due to its broad application, the DPP must not become a market entry barrier for companies or interfere with existing value chains. Technologies for the DPP must be affordable and accessible to companies, regardless of their economic and geographical focus. Recent research concurs that the usage of data spaces to manage the DPP system facilitates the secure exchange and easy linking of data. However, there is still no consensus on what information the DPP should contain. In this paper, we analyze the requirements that a DPP system must meet, discuss the role that the combination of knowledge graphs and data spaces could play, and outline their potential using the example of tracking the carbon footprint of products. The results create a basis that may considerably facilitate the future implementation of DPP systems.

ABSTRACT. Data spaces must be built bottom up, especially by leveraging trusted edge-based system architectures for processing data almost in real time closer to the data sources and enrich data/transform data for more added value, ensuring data privacy and security. More- over, there must be flexible context-driven access control models for managing data sharing for different applications in the data spaces ecosystems. Current IDS protocol does not provide a fully edge-based system architectures for data sharing. We aim to enable trusted edge-based data spaces that make the best out of Edge com- puting, but also are IDS-compliant. We propose an IDS-compliant framework that enables dynamic, context-driven, Edge-based IoT data sharing as a service (IDS4Edge). Our IDS4Edge framework can enable context-driven IoT data sharing by implementing flexible access control policies on top of IDS-connectors tailored to specific application-level IoT contexts. We have implemented a proof-of- concept to demonstrate our proposed solution. Our Edge-based IDS4Edge framework dynamically enforces tenant-specific access control policies on shared data, adapting in real-time to changes in IoT contexts and contractual agreements at the Edge.

ABSTRACT. This paper introduces a vision for Quantum Software Development lifecycle, proposing a hybrid full-stack iterative model that integrates quantum and classical computing. Addressing the current challenges in Quantum Computing (QC) such as the need for integrating diverse programming languages and managing the complexities of quantum-classical systems, this model is rooted in the principles of DevOps and continuous software engineering. It presents a comprehensive lifecycle for quantum software development, encompassing quantum-agnostic coding, testing, deployment, cloud computing services, orchestration, translation, execution, and interpretation phases. Each phase is designed to accommodate the unique demands of QC, enabling traditional software developers to engage with QC environments without needing in-depth QC expertise. The paper presents a detailed implementation roadmap, utilizing a range of existing tools and frameworks, thereby making quantum software development more accessible and efficient. The proposed model not only addresses current challenges in quantum software development but also makes a substantial contribution to the field of Quantum Software Engineering (QSE). By proposing a structured and accessible model, it sets the stage for further advancements and research in QSE, enhancing its practicality and relevance in a wide range of applications.

ABSTRACT. This paper reviews a quantum method for k-means clustering on NISQ computers and proposes a method to improve its accuracy. In addition, a quantum k-means clustering algorithm that efficiently utilizes qubit resources to calculate multiple distances simultaneously is introduced. Experiments on both quantum simulator and real quantum systems were conducted to compare the performance of the method proposed in this paper with the previous method. The results show that the proposed method achieves higher accuracy and efficiency in quantum simulator and even greater improvements in real quantum systems.

ABSTRACT. Connected autonomous vehicles (CAVs) require extensive data processing to make real-time decisions such as obstacle detection, collision avoidance, and route optimization. To enable fast and accurate decision-making in CAVs, services traditionally executed in cloud data centers are being moved to the edge to minimize data transfer from vehicles to the cloud. We present a framework called Q-Edge that utilizes quantum computing at the network edge to improve the speed and effectiveness of data analysis for CAVs. By integrating quantum computing principles such as superposition, entanglement, and teleportation with edge computing, we propose a solution to address the latency and bandwidth issues inherent in traditional cloud computing methods. This approach, leveraging quantum software engineering practices, enables real-time decision-making, optimizing CAV performance and enhancing traffic management, urban mobility, and road safety. Our results demonstrate that the proposed method efficiently performs data analytics that positively impact traffic management, urban mobility, and road safety.

ABSTRACT. This study aimed to develop a language model for the early detection of grooming in Korean. Based on PAN12 Korean dataset, we conducted early detection experiments using BERT-based models and Large Language Models (LLMs). In place of the window method, which references consecutive previous sentences, we introduced a memory method that references previous sentences similar to the input sentence and confirmed that the memory method outperforms the window method. We used reference sizes of 3, 5, and 10 sentences for each conversation. As the number of previous sentences increased, the memory method showed improved performance. We evaluated performance using F1 score, accuracy, speed, and latency-weighted F1. To address the limitations of latency-weighted F1, we introduced a new metric, Human-Model-Ratio (HMR).

ABSTRACT. In today's digital landscape, collaborative tools are critical for virtual teamwork, with comments as a key mechanism for communication and feedback. Our project, within the Natural Language Processing (NLP) domain, focuses on improving comment handling in collaborative environments using advanced machine learning methods. We developed a triage system that categorizes and prioritizes comments to help us efficiently address the most critical feedback. Building on previous work, we employed transformer models like BERT and RoBERTa, which showed strong performance in classifying comments when fine-tuned on our dataset. To enhance the handling of hierarchical structures, we experimented with Hierarchical Capsule Networks (HcapsNet) and Hierarchical Attention Networks (HAN). Additionally, GEMMA-2B, a large language model, demonstrated strong results in F1-score and precision while providing zero-shot and few-shot learning capabilities. The framework classifies and prioritizes comments based on six dimensions: urgency, importance, sentiment, actionability, resolution status, and thematic relevance. It incorporates rule-based logic alongside pre-trained NLP models, including GEMMA-2B for intent classification, Hugging Face models for sentiment analysis, and Latent Dirichlet Allocation (LDA) for topic modeling. This approach supports the efficient management of comments by prioritizing those that require immediate attention and improving the collaborative process.

ABSTRACT. The biodiversity of tropical environments offers a rich variety of species for the process of finding new drugs based on Natural Products. Databases like The Brazilian Biodiversity Natural Products Database (NUBBE$_{DB}$), where they hold compounds and characteristics about them, are important for computational assistance. However, these databases are difficult to update since data about compounds is mostly published in academic papers. Therefore, automatic Knowledge Extraction like on the state-of-the-art Benchmark for Natural Product Knowledge Extraction from Academic Literature (NatUKE), is an important task for the field. The dataset uses a Knowledge Graph version of the NUBBE$_{DB}$ and it evaluates different Knowledge Graph Embedding models for the task. The best performer from NatUKE is an embedding propagation model that uses pre-trained language models as the start-up embedding for the nodes that contain text data. This work investigates two avenues for increasing performance out of NatUKE. We focused on better text extraction from PDFs and using Large Language Models as the start-up embeddings. Our results surpassed state-of-the-art in 3 out of 5 extracted features while maintaining competitive performance on the remaining features.

ABSTRACT. Fallacies are arguments that seem valid but contain logical flaws. During the COVID-19 pandemic, they played a role in spreading misinformation, causing confusion and eroding public trust in health measures. Therefore, there is a critical need for automated tools to identify fallacies in media, which can help mitigate harmful narratives in future health crises. We present two key contributions to address this task. First, we introduce FALCON, a multi-label, graph-based dataset containing COVID-19-related tweets. This dataset includes expert annotations for six fallacy types—loaded language, appeal to fear, appeal to ridicule, hasty generalization, ad hominem, and false dilemma—and allows for the detection of multiple fallacies in a single tweet. The dataset’s graph structure enables analysis of the relationships between fallacies and their progression in conversations. Second, we evaluate the performance of language models on this dataset and propose a dual-transformer architecture that integrates engineered features. Beyond model ranking, we conduct statistical analyses to assess the impact of individual features on model performance.

ABSTRACT. This study investigates the use of Large Language Models (LLMs), specifically ChatGPT-4o, to enhance the moderation of online deliberative processes. Traditionally, decision-making has been controlled by small groups, often excluding the vital insights that crowd intelligence can provide. As global challenges grow more complex, broader and more inclusive participation is essential. While online platforms allow for such large-scale participation, they face significant issues, including content fragmentation, low signal-to-noise ratios, and inefficient argumentation. Human moderators can address these challenges, but the cost of employing them at scale is prohibitively expensive. This research introduces a more scalable solution by leveraging LLMs to automate critical moderation tasks, including unbundling multiple ideas, categorizing them into solutions, metrics, and barriers, and implementing efficient argument mining and classification techniques. Additionally, it evaluates the effectiveness of different prompting styles in optimizing moderation. The findings demonstrate that LLMs can successfully moderate key aspects of large-scale online deliberations, such as unbundling and categorization, improving the structure of discussions and representing a significant step forward in collective decision-making.

ABSTRACT. LogSage (Log Summarization Assistant with Guided Enhancement) is an innovative system designed to automate the generation of summaries from log data by applying few-shot learning and Reinforcement Learning with Human Feedback (RLHF) techniques. The system’s architecture incorporates a three-stage process. Initially, LogSage is trained using pairs of log entries and their corresponding human-written summaries to establish foundational understanding. Then, a base model processes these log entries, extracting essential content for summarization. To ensure the quality of the generated summaries, a reward model evaluates the outputs using expert human feedback. This feedback is instrumental in calibrating the summaries’ accuracy and relevancy. Finally, an aligned model refines the summaries based on the reward assessments, producing outputs that closely emulate human-quality text. LogSage represents a significant advancement in log analysis technology, offering precise, scalable, and efficient summarization capabilities that enhance both interpretability and actionable insights from voluminous log data.

ABSTRACT. This paper presents C2|Q>, a quantum software framework designed to bridge the gap between classical and quantum programming, specifically aimed at traditional developers who lack quantum computing expertise but want to leverage the potential benefits of quantum computers. C2|Q> transforms classical code written in high-level languages into quantum representations, enabling execution on Noisy Intermediate-Scale Quantum (NISQ) computers or simulators, and provides a user-friendly interface. The framework focuses on problems where quantum computing offers advantages, particularly nondeterministic polynomial-time complete (NP-complete) problems, which are considered classically hard. The key components of C2|Q> include a parser that identifies problem categories, extracts inputs, and converts classical inputs into quantum formats; a reducer that reformulates NP problems into satisfiability problems (SAT) when necessary; and an algorithm generator that constructs corresponding Hamiltonians and recommends feasible quantum algorithms. Additionally, the C2|Q> backend includes modules to optimize and compile quantum circuits, request and select suitable quantum hardware, and interpret computational results. The proposed framework empowers researchers and practitioners without prior quantum experience to harness the potential of hybrid quantum computation.

ABSTRACT. Recent advancements in Artificial Intelligence (AI) and more recently, Generative AI (GenAI) approaches have introduced both new challenges and opportunities in educational settings; yet, its effect on rural schools, which already face educational inequities, remains unclear. This study employs a mixed-methods approach using surveys and interviews to explore the current and potential roles of GenAI in rural high school classrooms across three U.S. regions. Preliminary findings reveal mixed perceptions: rural teachers value GenAI’s ability to personalize learning but worry about encountering misinformation and feel unprepared to mitigate these risks due to their limited AI literacy. While GenAI offers potential to enhance students’ tech skills and reduce resource disparities, barriers like unreliable internet access and a lack of students owning personal devices still hinder its effectiveness, leaving both teachers and students under-supported in fully leveraging the technology. Overall, this study aims to explore rural teachers’ experiences with GenAI to help develop strategies for fair and effective integration that address their unique challenges.

ABSTRACT. In my PhD, I will design, implement, and test an innovative approach supporting teachers in the generation of case studies for the simulation of computer-interpretable clinical guidelines. Constraint Programming will be exploited as a methodology to conciliate the different knowledge sources involved in the generation and produce case studies following the teachers’ instruction. The approach will be tested in an educational project involving a medical course at my university.

ABSTRACT. Precision agriculture faces significant challenges in optimizing water usage while maintaining crop yields, particularly as climate change intensifies water scarcity. This paper introduces Dynamic Adversarial Soil Modeling (DASM), an innovative framework that synergistically integrates Generative Adversarial Networks (GANs) and Confident Learning (CL) for precision irrigation. DASM’s novel architecture comprises: (a) a conditional GAN for synthetic soil state generation, capturing complex interdependencies between soil properties; (b) a multi-class CL algorithm for robust data quality assessment of both sensor and synthetic data; and (c) a multi-task learning optimization module for holistic ecosystem modeling. The system employs transfer learning for efficient cross-domain adaptation and outputs probabilistic soil state distributions, enabling risk-aware decision-making. Compared to traditional fixed-interval irrigation systems, DASM achieved a significant reduction in water usage while increasing crop yields by 12-16%.

ABSTRACT. In this article, we present a novel approach to program analysis through selective concrete execution. While static analysis of ELF binaries is necessarily limited by the theoretical undecidability of control flow and data flow analysis algorithms, we detail a new approach to reverse engineering through selective concrete execution of arbitrary functions within a x86_64 GNU/Linux binary by transforming ELF applications into shared libraries. This approach, named Procedural Debugging, allows us to empirically recover information about function parameters and return values without resorting to any disassembly or decompilation, which are undecidable in general. In turn, this dynamic approach may be used as a feedback loop into existing program analyzers, being them static, fuzzing, symbolic, or concolic, to enrich their understanding of application interfaces. We publish an open-source framework, named the Witchcraft Compiler Collection, under a permissive MIT/BSD license, implementing binary libification, procedural debugging, and automatic function prototype annotations with the hope of benefiting the security community.

ABSTRACT. In this project, we introduce Cross-Sectional Adaptive Transfer Learning (CATL), a novel transfer learning algorithm for signal classification in high-energy physics. CATL estimates a representation-based task distance between source and target background tasks using event cross-sections. We demonstrate the validity of CATL, its ability to produce robust classifications, and its efficacy in minimizing empirical loss for signal tasks. We then apply CATL to the ongoing search for the dark photon to compare results against previous methodologies. These results are generalized by producing risk bounds through gradient descent. Extensive model validation across multiple datasets, including cross-validation and statistical tests, provides strong evidence of the model's robustness and adaptability. The CATL methodology consistently yields improvements in signal efficiency by over 20% while simultaneously increasing target-aware background rejection compared to standard signal classifiers.

ABSTRACT. This paper examines computational social models in decentralized applications, focusing on Pump.fun, a platform for real-time trading of meme-based tokens on the Solana blockchain. The study explored the link between viral social trends on Twitter and the performance of meme tokens on Pump.fun. Surprisingly, tokens related to mainstream events had limited success, with only 289 out of 986,105 meme tokens launched between July and September 2024 tied to significant global events. Instead, subculture-driven memes, especially those about niche topics like cats, dogs, and internet personas such as Giga Chad, dominated, representing 85% of successful tokens. A new metric, the “Meme Pump dot fun metric” (MPDFM), was developed to predict the likelihood of tokens exceeding a market cap of $69,000 within 24 hours of launch. Validated using the Jaccard similarity index, this metric showed a high similarity (0.82) between top tokens and successful tokens from a future date. This shift highlights the influence of internet-native communities over global events in decentralized financial ecosystems.

ABSTRACT. Deep learning models often perform suboptimally compared to tree-based models for tabular data. Moreover, the application of vision models to molecular sequences is relatively unexplored in the literature. This research introduces a novel method for transforming molecular sequences into images to enable sequence classification using vision-based deep learning models. Utilizing Rips complex construction in combination with Chaos Game Representation (CGR), we map molecular sequence elements to unique coordinates in a two-dimensional space, compute the distance matrix, and construct the Rips complex for each sequence. This approach effectively captures the structural and topological features of molecular sequences, converting them into informative visual representations suitable for deep learning classifiers. Applied to the different anti-cancer peptide sequence datasets as the case study, our method achieves higher predictive performance compared to existing methods. The results demonstrate that the sequence-to-image transformation retains critical information necessary for accurate classification, providing a powerful tool for bioinformatics and computational biology applications. This framework offers a scalable and robust solution for leveraging vision deep learning models in the analysis and classification of diverse molecular sequences. This framework not only enhances bioinformatics and computational biology applications but also opens new avenues for the development of health recommender systems, particularly in areas requiring precise molecular sequence analysis, such as personalized medicine and drug discovery.

ABSTRACT. With the surge in audio data available today, there is a growing need for effective dialogue summarization. This study conducts two experiments using two LLMs, BART and Mistral, to assess dialogue summarization. The first experiment evaluates model performance, while the second examines the impact of upstream errors from Automatic Speech Recognition (ASR) and Machine Translation (MT) on summarization performance. Results indicate that SummaC, a commonly used evaluation metric, is unreliable for dialogue summarization. Additionally, Mistral's summarization performance is more sensitive to upstream errors than BART's.

ABSTRACT. Graph Structure Learning (GSL) techniques improve real-world graph representations, enabling Graph Neural Networks (GNNs) to be applied to unstructured domains like 3D face analysis. GSL dynamically learns connection weights within message-passing layers of a GNN, particularly in Graph Convolutional Networks (GCNs). This becomes crucial when working with small sample datasets, where methods like Transformers, which require large data for training, are less effective. In this paper, we introduce a kernel-attentive Graph Convolutional Network (KA-GCN) designed to integrate a positional bias through attention mechanisms into small-scale 3D face analysis tasks. This approach combines kernel- and attention-based mechanisms to adjust different distance measures and learn the adjacency matrix. Extensive experiments on the FaceScape public dataset and a private dataset featuring facial dysmorphisms show that our method outperforms state-of-the-art models in both effectiveness and robustness. The code is available at \url{https://anonymous.4open.science/r/KA-CONV-6490}.

ABSTRACT. The state-of-the-art approaches to Human Activity Recognition rely on deep learning models to extract complex translational-invariant features (CNNs) or to exploit the temporal dependencies (LSTMs) from sensors’ time series data. However, there are also other dependencies between sensors beyond the time dimension, e.g. physical proximity, which are equally important for characterization of human activities. In this work, we leverage such spatial dependencies by modeling them as a graph. Using Graph Neural Networks (GNNs), we learn global and local representations of the intra- and inter-sensor dependencies. We empirically show that by maximizing the mutual information between the local and global representations, the performance of the recognition models can be significantly improved. Our results show a clear improvement over previous works based on CNNs, LSTMs, Attention-based and other more complex GNNs-based architectures. Source code: https://anonymous.4open.science/r/GNNs4HAR-31C2

ABSTRACT. Merging Internet (web2) identities with blockchain (web3) identities is increasingly important for enhancing user experience and ensuring regulatory compliance. However, conventional solutions that map web2 identities to web3 accounts often lead to privacy concerns and fragmented identifiers across networks. To address these challenges, we propose a new identity scheme named Address Abstraction (AA), which redefines blockchain address and signing systems while preserving key properties: uniqueness, immutability, and privacy-preservation. This approach eliminates the limitations of chain-specific identity systems, enabling users to interact with multiple blockchains using their web2 certificates and unified identifiers. This chain-agnostic identifier also promotes cross-chain compatibility. We further present Zero-Knowledge Address Abstraction (zkAA), an implementation of AA that uses zero-knowledge proofs to uphold AA’s core properties. Additionally, a proof aggregation technique combines multiple proofs into one, achieving approximately 5.5 times gas cost savings during verification in real-world scenarios. As of August 2024, zkAA with proof aggregation incurs an additional cost of only $0.66 per transaction on Ethereum.

ABSTRACT. Tornado Cash is a decentralized application (dApp) that runs on Ethereum Virtual Machine (EVM) compatible networks to enhance users’ privacy in terms of user transaction history over the blockchain. This dApp achieves its goal by enabling users to deposit currencies into designated pools and subsequently withdraw them, severing the link between depositor and withdrawer addresses. At deposit time, Tornado Cash communicates to users the level of privacy they will benefit from (anonymity set) by depositing currencies into one of its pools. Existing analyses have indicated discrepancies between the claimed anonymity set and the actual level of privacy provided, primarily attributed to users’ incorrect utilization of the dApp. This paper explores the road towards a new way to challenge the dApp’s proposed anonymity set by examining wallet fingerprints, a factor not directly related to user behavior within the application. The findings of this research shed light on the potential for creating links between clusters of users in TC according to the new proposed approach and raise a privacy concern within the Ethereum network, resulting in 13203 transactions, over 66948, linkable to the wallet used to initialize them.

ABSTRACT. In a world becoming increasingly digital for all types of interactions, from shopping to communicating, many people still face accessibility challenges. In particular, deafblind individuals, i.e. both visually and auditory impaired, rely mostly on tactile means to access information, of which most common interfaces are devoided. Researchers have investigated the development of alternative communication tools, through wearables, specific devices or simple coded vibrations on smartphones. Yet, these usually require an additional device or have not been fully evaluated with users. Thus, we propose to leverage a multitouch localised tactile feedback technology, which can be integrated into a surface, such as a tablet or smartphone. We have adapted the Lorm alphabet using the prototype capabilities to render a range of stimuli, including taps, slides and vibrotactile patterns. A first evaluation of the alphabet with 18 sighted participants validated the principle with a high letter recognition rate of 90%.

ABSTRACT. Road anomaly detection is crucial for improving vehicle safety and comfort by identifying irregularities that impact the driving experience. This study aims to develop a multi-purpose application that creates a comprehensive data layer about road conditions. Such a data layer could integrate with applications like Google Maps, allowing users to choose routes based on real-time road condition information. This is particularly beneficial for users with specific mobility needs. The application uses sensor data from smartphones. By using threshold-based methods to analyze accelerometer data, coupled with a dynamic rolling window approach that adjusts based on local speed, this system offers robust detection of road anomalies. This method has been validated against ground truth data, demonstrating its potential to provide reliable information to many users.

ABSTRACT. This paper presents an approach to enhancing Augmentative and Alternative Communication (AAC) systems tailored to Brazilian Portugues by integrating Colourful Semantics (CS) with transformer-based language models. We introduce an adapted BERT model, BERTptCS, which incorporates the CS framework for improved prediction of communication cards. The primary aim is to enhance the accuracy and contextual relevance of communication card predictions, which are essential in AAC systems for individuals with complex communication needs (CCN). We compared BERTptCS with a baseline model, BERTptAAC, which lacks CS integration. Our results demonstrate that BERTptCS significantly outperforms BERTptAAC in various metrics, including top-k accuracy, Mean Reciprocal Rank (MRR), and Entropy@K. Integrating CS into the language model improves prediction accuracy and offers a more intuitive and contextual understanding of user inputs, facilitating more effective communication.

ABSTRACT. Recommender systems (RSs) are extensively utilized in e-commerce to predict users' future preferences for unseen items based on historical user-item interactions. These systems, however, are vulnerable to manipulations by malicious actors, such as unsolicited users or vendors, through various shilling attacks. Such attacks intentionally skew recommendations by injecting biased data to promote or demote certain products or services. To address this issue, we propose a generative model called Diff-WGAN, designed to mitigate the impact of shilling attacks within an adversarial learning framework. Diff-WGAN uses a combination of Diffusion model (DiffRec) and a GAN framework (CFGAN) to leverage the adversarial advantages of GAN and the personalization advantages of DiffRec. We employ a diffusion model as the generator to process the inherently noisy and sparse historical user-item interactions. The discriminator is a multi-layer perceptron that employs Wasserstein distance as its loss function. We conducted preliminary experiments using four well-known evaluation datasets: MovieLens 100K, MovieLens 1M, Amazon-apps, and Yelp. By simulating various attack scenarios by integrating fake interactions in the dataset, we demonstrate that our Diff-WGAN model outperforms baseline models across most datasets and attack types, showing better resistance against shilling attacks.

ABSTRACT. With the vast amount of reviews available in the Web, in the past decades, a growing share of literature has focused on sentiment analysis. Aspect-based sentiment classification is the subtask that seeks to detect the sentiment expressed by the content creators towards a defined target within a sentence. This paper introduces two novel unsupervised attentional neural network models for aspect-based sentiment classification, and tests them on English restaurant reviews. The first model employs an autoencoder-like structure to learn a sentiment embedding matrix where each row of the matrix represents the embedding for one sentiment. To improve the model, a target-based attention mechanism is included that de-emphasizes irrelevant words. Last, a redundancy and a seed regularization term constrain the sentiment embedding matrix. The second model extends the first by including a Bi-LSTM layer in the attention mechanism to exploit contextual information. Although both models construct meaningful sentiment embeddings, experimental results indicate that the inclusion of the Bi-LSTM in the attention mechanism leads to a more precise attention mechanisms and, thus, better predictions. The best model, i.e., the second, outperforms all investigated unsupervised and weakly supervised algorithms for aspect-based sentiment classification from the literature.

ABSTRACT. Reinforcement learning (RL) policies can exhibit unsafe behaviors and are often difficult to explain. While local explainability methods for RL offer insights into specific decisions, they often lack the temporal context for comprehensive explanations, especially for understanding safe behavior. This paper combines local explainable RL with PCTL model checking to explain complex, safe sequential decision-making over time, providing deeper insights into the trained RL policy. Our method uses five inputs: (1) a Markov Decision Process (MDP) representing the RL environment, (2) a trained policy, (3) a PCTL formula for safety assessment, (4) a local explainable RL method, and (5) a PCTL formula for the explainable RL method quantification at each state. By incrementally building reachable parts of the MDP guided by the trained policy and incorporating additional explainable features into the MDP's factored state representation, we verify the policy's safety in an explainable manner using PCTL model checking over these explainable features. Through diverse RL environments and different PCTL queries, we demonstrate that our method can explain trained RL policies in the context of explainable safety.

ABSTRACT. The increasing reliance on diffusion models for generating synthetic images has amplified concerns about the unauthorized use of personal data, particularly facial images, in model training. In this paper, we introduce a novel identity inference framework to hold model owners accountable for including individuals' identities in their training data. Our approach moves beyond traditional membership inference attacks by focusing on identity-level inference, providing a new perspective on data privacy violations. Through comprehensive evaluations on two facial image datasets, Labeled Faces in the Wild (LFW) and CelebA, our experiments demonstrate that the proposed membership inference attack surpasses baseline methods, achieving an attack success rate of up to 89\% and an AUC-ROC of 0.91, while the identity inference attack attains 92\% on LDM models trained on LFW, and the data extraction attack achieves 91.6\% accuracy on DDPMs, validating the effectiveness of our approach across diffusion models.

ABSTRACT. The article discusses a novel method for protecting privacy in text sentiment classification, a task commonly used in fields like social media and customer service. Traditional models often risk exposing sensitive information, such as user preferences or personal traits. To address this, we propose a "Text Sentiment-aware Privacy Protection Method" based on truncated exponent mechanisms. This approach combines sentiment-aware word embeddings with a differential privacy technique, enabling sentiment analysis with a balanced trade-off between privacy protection and classification accuracy. The method was validated in various scenarios, demonstrating its effectiveness in balancing privacy protection and data utility.

ABSTRACT. Convolutional Neural Networks (CNNs) have demonstrated considerable efficacy in classification tasks across several domains, including computer vision and healthcare. However, these networks are vulnerable to adversarial attacks, which introduce small perturbations in the input in order to cause misclassifications. While white-box adversarial attacks are not practically applicable in a real-world scenario, black-box attacks are generally non-deterministic and based on random methodologies. This work proposes a novel adversarial attack based on grey-box methodologies. First, it identifies the most important sections within the input image that contribute the most to the final classification, by leveraging information related to the architecture of the CNN. Finally, with a deterministic approach, the attack modifies the pixels of the aforementioned sections, introducing a perturbation of the smallest possible magnitude. The objective is to generate an adversarial sample that is indistinguishable from the original image. Extensive experimental results demonstrate that our attack has a high success rate, and that generate adversarial images that are visually indistinguishable from the original images.

ABSTRACT. Supervised Machine learning (ML) is used in many safety-critical applications, such as self-driving cars and medical imaging. Unfortunately, many training datasets have been discovered to contain faults. The accuracy of individual models when trained with faulty datasets can significantly degrade. In comparison, ensembles, consisting of multiple models combined through simple majority voting, are able to retain accuracy despite training data faults, due to their classification diversity, and are thus more resilient. However, there are many different ways to generate ML ensembles, and their accuracy can significantly differ. This creates a large search space for ensembles, making it challenging to find ensembles that maximize accuracy despite training data faults. We identify three different ways to generate diverse ML models, and present D-semble, a technique that uses Genetic Algorithms and diversity to efficiently search for resilient ensembles. We evaluate D-semble by measuring the balanced accuracies and F1-scores of ensembles it finds. Compared with bagging, greedy search, random selection, and the best individual model, ensembles found by D-semble are on average 9%, 16%, 28%, 32% more resilient respectively.