Technical debt prediction involves anticipating problems and challenges in software projects. Because of lousy coding choices, these projects are often far from the optimal design. These issues can accumulate over time and affect long-term maintainability and stability. Numerous studies are available for predicting technical debt. For instance, some studies utilize static analysis tools like Pylint(1), and others use dynamic code analysis tools like SZZ [1]. This study addresses AI predictions using data obtained from static analysis tools. This paper presents technical debt prediction using static code analysis data generated with SonarQube(2) and Radon(3). It covers both short-term and long-term predictions by analyzing code quality metrics such as complexity, duplication, and adherence to standards. The study also integrates git commit classification to analyze commit history, providing valuable insights into the evolution and management of technical debt within the project.
Automated machines are widely used in industrial environments for the production of different items and generate lots of data following these processes. The proper execution of the machines influences the production output, thus the detection of anomalous behavior in machines' activity using the generated data must be considered to avoid unpleasant outcomes. This paper presents a novel ensemble unsupervised anomaly detection method characterized by parameterized prediction. The proposed method consists of two stages - the first stage uses statistical-based methods to assign artificial labels to the input data. In the second stage, the artificially labeled instances and a feature bagging technique are employed to construct the model's estimators - each estimator calculates the percentiles of the distances computed between its centroid and each instance from the training subset. The model's prediction function is parameterized by a percentile rank. Each estimator computes the distance between its centroid and the evaluated instance: if this distance exceeds the value of the specified percentile, the estimator classifies the instance as an anomaly. A majority vote is applied to establish the outcome.
Lung cancer persists as a global leader in cancer-related deaths, highlighting the critical need for precise and efficient detection methods. This paper investigates the use of the Medical Segmentation Decathlon dataset to train neural networks for lung cancer segmentation in CT scans via semantic segmentation. We propose and evaluate four new data adaptation techniques specifically designed for this dataset, with each technique being assessed using U-Net-based architectures. Our approach incorporates a thorough exploratory data analysis to uncover the dataset's strengths and weaknesses, which in turn guided our data preprocessing and augmentation strategies.
Disassembly is the process that translates machine code into a higher-level and more human-readable form. From the security analysis and malware detection perspective, it is an essential technique. However, traditional disassembly tools often display the information in a textual format, thus making it difficult to navigate and understand. In this paper, we explore the concepts involved in creating a viewer for disassembly that focuses on static analysis and boosts the user experience like clearly displaying the code, showcasing relevant parts, guiding the user, offering an interactive experience, map data structures and seamless integration with other tools. Furthermore, we emphasize the importance of having both a user-friendly and also customizable interface without having to concede on overall speed and performance. In this manner, even a novice in security analysis could have a positive experience. Considering all these features, the final result leads to a more productive and efficient analysis.
Testing and validating embedded software is paramount when it comes to safety-critical systems. BTC EMBEDDEDPlatform((R)) supports these activities using various approaches including test case definition means. In this system description we present SCRIPTVIEW, a domain-specific language designed to support test case implementation in a simple programmatic way. The language has been integrated in our platform and can be used in production by our customers.
This study explores feature selection for classifying galaxy morphology using the extensive Galaxy Zoo 2 dataset. We investigate supervised and unsupervised learning methods to group galaxies based on key features, aiming to replicate supervised learning results. We evaluate various feature selection methods and compare them to an existing classification approach. Our results demonstrate that a reduced set of features based on adjusted vote fractions improves classification accuracy and potentially reduces computational complexity. While unsuper-vised clustering partially groups galaxies by morphology, further optimization is required. This work suggests that feature selection and unsupervised learning are promising techniques for the efficient classification of large galaxy datasets in upcoming astronomical surveys.
Serverless computing, also known as Function as a Service (FaaS), simplifies cloud application development by abstracting server management. However, cold start latency, which occurs when initializing the execution environment, poses sig-nificant challenges. This work proposes an Apache Open Whisk-integrated dynamic caching and request routing system to mit-igate cold start latency. By leveraging intelligent caching, the system aims to reduce response times for serverless functions. Comprehensive performance evaluations demonstrate that the proposed system reduced average response times by up to 7 times, while also ensuring efficient resource utilization. These findings provide practical insights and scalable solutions to enhance the efficiency and reliability of serverless architectures, particularly for latency-sensitive applications.
The proliferation of sophisticated AI generative models like GPT-4 has revolutionized natural language processing (NLP) but also raised critical concerns about content authenticity in academia, media, and digital communications. This paper introduces a dual-approach AI-generated text detector that leverages both traditional machine learning (ML) techniques and advanced fine-tuned large language models (LLMs). Utilizing a comprehensive dataset of over 350,000 samples from five benchmark sources, our approach demonstrated robust performance, with conventional ML methods achieving 91-92% accuracy (0.97 ROC-AUC) and fine-tuned LLMs such as BERT and RoBERTa reaching 97-98% accuracy (0.99 ROC-AUC). We developed TruAIText, a practical tool that integrates these models to provide detailed analysis of AI-generated content, including paragraph-level probabilities. Despite its efficacy, the tool requires ongoing updates to counteract adversarial manipulation.
Medical image segmentation is one of the most important topics in the field of computer vision and plays a crucial role in computer-aided diagnosis. U-Net paved the way for a series of variants that took advantage of the key characteristics of this network. In this article, several features proposed in different variants of U-Net are adapted and experimented upon to create a new architecture that maintains the idea of a U-shaped structure. The proposed architecture takes advantage of the efficient depth-wise separable convolution, but with a twist. Instead of using the pointwise convolution as the last step in the depth-wise separable convolution, it utilizes the so-called Ghost Module. This results in a highly efficient network with a reduced complexity, that still has excellent segmentation performance. We compared SUDS with U-Net and its variants across multiple segmentation tasks from two categories, skin lesion segmentation and colonscopy segmentation. Experiments demonstrate that SUDS has similar segmentation accuracy compared to the other networks, while the number of parameters and floating-point operations are greatly reduced.
Many theorems in mathematics have the form of an implication, an equivalence, or an equality, and in the standard prover in the Theorema system such formulas are handled by rewriting. Definitions of new function- or predicate symbols are yet another example of formulas that require rewriting in their treatment in the Theorema system. Both theorems and definitions in practice often carry conditions under which they are valid. Rewriting is, thus, only valid in cases where all side-conditions are met. On the other hand, many of these side-conditions are trivial and when presenting a proof we do not want to distract the reader with lengthy derivations that justify the side-conditions. The goal of this paper is to present the design and implementation of a mechanism that efficiently checks side-conditions in rewriting while preserving the readability and the explanatory power of a mathematical proof, which has always been of central interest in the development of the Theorema system.
There is currently a widespread interest in using phased execution models in the design of real-time systems, stimulated by their ability to reduce contention in accessing shared resources. The main beneficiary of such approaches is the wide category of Commercial Off-The-Shelf (COTS) multi-core systems, where complexity is an underlying issue. However, the advantage of predictability comes at a cost, which often resides in the additional decisions that are usually made when using phased execution models. One such decision is employing non-preemptive scheduling; as is well known, this leads to NP-hardness for most instances of the scheduling problem, while preemption can avoid this level of complexity in many cases. It is no surprise, then, that recent efforts have attempted to explore the possibilities offered by preemptive scheduling in combination with phased execution models. Unlike existing work, the current paper focuses on the scratchpad memory consumption in preemptive phased execution models. The starting point, the model introduced by the authors of a previous article, was meant for the general analysis of the efficiency of preemptive scheduling in phased execution models. Notably, two methods for handling intermediate data are introduced, namely, the Waiting Time Minimizing Preemption (WMPM) and the Overhead Minimizing Preemption (OMPM). The current paper analyses the requirements of each approach in terms of scratchpad memory. It also introduces a hybrid method, in order to achieve a finer balance between memory consumption and schedulability constraints.
Analysing malicious files is a lengthy task and, given a set of hundreds, even thousands of files, a security researcher would have to spend multiple hours looking at each sample. Considering that new malware is created every day, a way to reduce the analysis time is mandatory. This paper presents the Malware Analysis and Clustering Engine (MACE), a solution implemented to address this problem. It offers the user a modular framework implementing multiple feature extraction methods and clustering algorithms, which can provide a multitude of clustering configurations. The users can experiment with differ-ent configurations and choose one that best suits their scenario, obtaining a clustering result that allows them to analyse only a single sample from each cluster and draw a conclusion both for it and all its peers. This approach significantly accelerates the analysis process. The engine was tested on four different datasets, each trying to exemplify a scenario that is likely to be encountered in real situations. The results indicate that MACE does achieve its goal, providing multiple configurations for each dataset that reach both great accuracy and a number of clusters that is close to the one that was expected.
This paper aims to provide a comprehensive analysis of the benefits of employing GraphLIME (Local Interpretable Model Explanations for Graph Neural Networks) for reliable diabetes mellitus prediction. Our focus is on highlighting the advantages of integrating GraphLIME with a features attention-mechanism, compared to the standard pairing of deep learning neural networks with the original LIME explainability method. This system enabled us to develop an effective approach for identifying the most relevant features and applying the attention mechanism solely to those features. We conducted a detailed comparison of the performance metrics between the two approaches. By incorporating the attention mechanism, the model reached an accuracy of 92.6% in addressing the problem. The model's performance is thoroughly illustrated, with results further assessed using the Receiver Operating Characteristic (ROC) curve. Applying this technique to a dataset of 768 patients with or without diabetes mellitus, we enhanced the model's performance by over 18%.
In logistics and industrial operations, optimizing crane scheduling is crucial for efficiency and resource allocation. Given the dynamic nature of crane scheduling environments, influenced by factors such as fluctuating order volumes and varying crane availability, robust scheduling solutions are imperative. This paper introduces a rule-based approach designed to optimize crane scheduling within a simulated dynamic and uncertain environment, focusing on real-time responsiveness, efficient task allocation, and continuous operational efficiency. Our approach was developed while participating in the “Crane Scheduling” track of the “Dynamic Stacking Optimization in Uncertain Environments” (DynStack) competition, held with the Genetic and Evolutionary Computation Conference (GECCO) 2024. Our experimental results underscore the efficacy of our approach in achieving zero delivery errors, maximizing the handling of blocks, and minimizing parking durations, hereby promoting cost-effectiveness and environmental sustainability in crane operations. As of July 17, 2024, our approach has secured first place in the “Crane Scheduling” competition.
We propose a solution for handling robot abort commands, illustrated by a real-world scenario involving household kitchen robots. The robot uses planning to find sequences of actions that must be performed in order to gracefully cancel a previously received command. The Planning Domain Definition Language (PDDL) is used to write a domain to model kitchen activities and behaviours, and this domain is enriched with knowledge from online ontologies and knowledge graphs, like DBPedia. We discuss the results obtained in different scenarios.
Phishing emails are among today's most common attack vectors since most enterprises and consumer users still rely on email for day-to-day operations. At the same time, some functionalities (such as MACRO support and Visual Basic Programming) available in Microsoft Office Suite make it easier for an attacker to inject malicious code into a document that can be sent via email. From a research perspective, analyzing such an attack implies understanding not only the artifacts that compose the kill chain steps but also being able to extract, from a document, the embedded macros and the scripts. This paper focuses on building such a system using the forensics platform GView by adding various plugins and functionalities designed to augment the existing support and allow a security researcher to analyze the content of an email and attached documents quickly.
In this paper, we focus on the monomial prediction problem in two settings: (1) Decide whether a particular monomial m is present in a composite function f := f(r) circle f(r-1) circle.... f(0), where f(i) are quadratic boolean functions, (2) Decide whether a particular monomial m is present in a composite function f := f(r) circle f(r-1).... f(0), where polynomials f(i) are efficiently computable by Probabilistic Generating circuits over rationals. Probabilistic generating circuits (PGCs) are economical representations of multivariate probability generating polynomials (PGPs), which capture many tractable probabilistic models in machine learning. The first problem has a strong connection with the security of symmetric-key primitives. Dinur and Shamir proposed the cube attack for distinguishing a cryptographic primitive from a random function, which can be thought of as an efficient monomial prediction. In a general setting, over any large finite field or integers, monomial prediction is known to be NP-hard. Here, we show that in the quadratic setting, the problem is circle plus P-complete. circle plus P is an interesting complexity class that is not known to contain NP, however, it is believed to contain computationally hard problems. On the other hand, we also present several new zero-sum distinguishers for 5-round Ascon, which is one of the ten finalists for NIST light weight cryptography standardization competition. We show that the second problem is #P-complete. It is known that PGCs have efficient inference, i.e. given a monomial, one can efficiently output (which signifies the probability) its coefficient in the polynomial computed by the circuit. However, a composition of such functions makes the inference hard. Composition of probabilistic models and their efficient inference play a crucial role in the semantic contextualization and framework of uncertainty theories in graphical modelling.
We investigate the abstractness of a continuation semantics for a calculus inspired by DNA computing. This semantic investigation is given in the framework of complete metric spaces, and uses the weak abstractness criterion introduced by us in recent work. We prove that the denotational semantics designed with continuations is weakly abstract with respect to the operational semantics of our calculus which involves multiparty synchronization. We show that the expected concurrency laws are satisfied in this semantics for the calculus under investigation.
This paper contains a comparison between a Genetic Algorithm (GA) and a Non-dominated Sorting Genetic Algorithm II (NSGA-II) on the Portfolio Optimisation Problem, based on the Modern Portfolio Theory proposed by Markowitz (1952, 1956). We use the real-world data of the S&P 500 index (quarterly returns of top 200 stocks from 2019 to 2024, and top 442 stocks from 2004 to 2016), and show that both algorithms can yield returns above the index, in mixed bull and bear market conditions. N-point Crossover accelerates algorithm convergence, and by using the Sharpe Ratio, the NSGA-II outperformed most models based on Stochastic Dominance we tested, according to the metrics in Bruni el. al. (2017).
The field of machine learning has increasingly focused on incremental learning, enabling systems to continually adapt and improve by integrating new knowledge while retaining previously learned information. One particular area of interest is class-incremental learning, where the learning system sequentially acquires new classes without access to or with limited exposure to past data. The primary challenge in class-incremental learning is catastrophic forgetting, wherein the model tends to overlook previously learned classes when confronted with new tasks. One of the class-incremental learning approaches to mitigate catastrophic forgetting is the Incremental Classifier and Representation Learning (iCaRL) framework. In this paper, we propose three new selection criteria for the iCaRL approach. Our best selection criterion, which uses the K-Means clustering algorithm to create diverse groups and then selects exemplars close to the centroids of the clusters, outperforms the original iCaRL selection criterion by over 16% for the MNIST dataset and by over 12% for the FashionMNIST dataset in terms of average accuracy. The full implementation of the iCaRL approach, along with the three proposed selection criteria and detailed experimental results logs, can be found in our publicly available GitHub repository. By contributing to the ongoing development of class-incremental learning techniques, we aim to support the creation of more effective and robust lifelong machine learning systems.
