arXiv subjects

Foutse Khomh

Publications and source records attributed to Foutse Khomh.

At least 55 records · Page 3Linked to original sources

arXiv2026

Improving the Robustness of Large Language Models for Code Tasks via Fine-tuning with Perturbed Data

Context: In the fast-paced evolution of software development, Large Language Models (LLMs) have become indispensable tools for tasks such as code generation, completion, analysis, and bug fixing. Ensuring the robustness of these models against potential vulnerabilities from handling diverse inputs is critical, as variations in input can lead to incorrect or insecure code outputs. Objective: This work aims to improve the robustness of LLMs for coding-related tasks against potential adversarial inputs. Specifically, we investigate how fine-tuning LLMs with perturbed datasets impacts their robustness against input perturbations. Method: We systematically evaluated LLM robustness by fine-tuning models using datasets perturbed at character-level, word-level, and sentence-level, comparing results against base models and models fine-tuned on unperturbed datasets. Results: Fine-tuning LLMs with perturbed datasets significantly improves model robustness (RD usually drops around 4\% - 6\%), especially for models with relatively weak robustness. However, this fine-tuning process typically results in a slight performance decrease (pass@1 usually drops around 1\% - 3\%) compared to fine-tuning with unperturbed datasets, although occasional performance improvements are observed. Conclusion \& Implications: Fine-tuning LLMs for coding tasks with perturbed data effectively enhances their robustness at the cost of a minor performance reduction, emphasizing the importance of balancing the robustness and performance of LLMs for coding applications.

cs.SE↗

arXiv2026

A Context-Driven Approach for Co-Auditing Smart Contracts with The Support of GPT-4 code interpreter

The surge in the adoption of smart contracts necessitates rigorous auditing to ensure their security and reliability. Manual auditing, although comprehensive, is time-consuming and heavily reliant on the auditor's expertise. With the rise of Large Language Models (LLMs), there is growing interest in leveraging them to assist auditors in the auditing process (co-auditing). However, the effectiveness of LLMs in smart contract co-auditing is contingent upon the design of the input prompts, especially in terms of context description and code length. This paper introduces a novel context-driven prompting technique for smart contract co-auditing. Our approach employs three techniques for context scoping and augmentation, encompassing code scoping to chunk long code into self-contained code segments based on code inter-dependencies, assessment scoping to enhance context description based on the target assessment goal, thereby limiting the search space, and reporting scoping to force a specific format for the generated response. Through empirical evaluations on publicly available vulnerable contracts, our method demonstrated a detection rate of 96\% for vulnerable functions, outperforming the native prompting approach, which detected only 53\%. To assess the reliability of our prompting approach, manual analysis of the results was conducted by expert auditors from our partner, Quantstamp, a world-leading smart contract auditing company. The experts' analysis indicates that, in unlabeled datasets, our proposed approach enhances the proficiency of the GPT-4 code interpreter in detecting vulnerabilities.

cs.SE↗

arXiv2026

Tri-LLM Cooperative Federated Zero-Shot Intrusion Detection with Semantic Disagreement and Trust-Aware Aggregation

Federated learning (FL) has become an effective paradigm for privacy-preserving, distributed Intrusion Detection Systems (IDS) in cyber-physical and Internet of Things (IoT) networks, where centralized data aggregation is often infeasible due to privacy and bandwidth constraints. Despite its advantages, most existing FL-based IDS assume closed-set learning and lack mechanisms such as uncertainty estimation, semantic generalization, and explicit modeling of epistemic ambiguity in zero-day attack scenarios. Additionally, robustness to heterogeneous and unreliable clients remains a challenge in practical applications. This paper introduces a semantics-driven federated IDS framework that incorporates language-derived semantic supervision into federated optimization, enabling open-set and zero-shot intrusion detection for previously unseen attack behaviors. The approach constructs semantic attack prototypes using a Tri-LLM ensemble of GPT-4o, DeepSeek-V3, and LLaMA-3-8B, aligning distributed telemetry features with high-level attack concepts. Inter-LLM semantic disagreement is modeled as epistemic uncertainty for zero-day risk estimation, while a trust-aware aggregation mechanism dynamically weights client updates based on reliability. Experimental results show stable semantic alignment across heterogeneous clients and consistent convergence. The framework achieves over 80% zero-shot detection accuracy on unseen attack patterns, improving zero-day discrimination by more than 10% compared to similarity-based baselines, while maintaining low aggregation instability in the presence of unreliable or compromised clients.

cs.CR↗

arXiv2026

Evaluating Implicit Regulatory Compliance in LLM Tool Invocation via Logic-Guided Synthesis

The integration of large language models (LLMs) into autonomous agents has enabled complex tool use, yet in high-stakes domains, these systems must strictly adhere to regulatory standards beyond simple functional correctness. However, existing benchmarks often overlook implicit regulatory compliance, thus failing to evaluate whether LLMs can autonomously enforce mandatory safety constraints. To fill this gap, we introduce LogiSafetyGen, a framework that converts unstructured regulations into Linear Temporal Logic oracles and employs logic-guided fuzzing to synthesize valid, safety-critical traces. Building on this framework, we construct LogiSafetyBench, a benchmark comprising 240 human-verified tasks that require LLMs to generate Python programs that satisfy both functional objectives and latent compliance rules. Evaluations of 13 state-of-the-art (SOTA) LLMs reveal that larger models, despite achieving better functional correctness, frequently prioritize task completion over safety, which results in non-compliant behavior.

cs.CL↗

arXiv2026

Empirical Characterization of Logging Smells in Machine Learning Code

\underline{Context:} Logging is a fundamental yet complex practice in software engineering, essential for monitoring, debugging, and auditing software systems. With the increasing integration of machine learning (ML) components into software systems, effective logging has become critical to ensure reproducibility, traceability, and observability throughout model training and deployment. Although various general-purpose and ML-specific logging frameworks exist, little is known about how these tools are actually used in practice or whether ML practitioners adopt consistent and effective logging strategies. To date, no empirical study has systematically characterized recurring bad logging practices--or logging smells--in ML System. \underline{Goal:} This study aims to empirically identify and characterize logging smells in ML systems, providing an evidence-based understanding of how logging is implemented and challenged in practice. \underline{Method:} We propose to conduct a large-scale mining of open-source ML repositories hosted on GitHub to catalogue recurring logging smells. Subsequently, a practitioner survey involving ML engineers will be conducted to assess the perceived relevance, severity, and frequency of the identified smells. \underline{Limitations:} % While The study's limitations include that While our findings may not be generalizable to closed-source industrial projects, we believe our study provides an essential step toward understanding and improving logging practices in ML development.

cs.SE↗

arXiv2026

An Empirical Study of Policy-as-Code Adoption in Open-Source Software Projects

\textbf{Context:} Policy-as-Code (PaC) has become a foundational approach for embedding governance, compliance, and security requirements directly into software systems. While organizations increasingly adopt PaC tools, the software engineering community lacks an empirical understanding of how these tools are used in real-world development practices. \textbf{Objective:} This paper aims to bridge this gap by conducting the first large-scale study of PaC usage in open-source software. Our goal is to characterize how PaC tools are adopted, what purposes they serve, and what governance activities they support across diverse software ecosystems. \textbf{Method:} We analyzed 399 GitHub repositories using nine widely adopted PaC tools. Our mixed-methods approach combines quantitative analysis of tool usage and project characteristics with a qualitative investigation of policy files. We further employ a Large Language Model (LLM)--assisted classification pipeline, refined through expert validation, to derive a taxonomy of PaC usage consisting of 5 categories and 15 sub-categories. \textbf{Results:} Our study reveals substantial diversity in PaC adoption. PaC tools are frequently used in early-stage projects and are heavily oriented toward governance, configuration control, and documentation. We also observe emerging PaC usage in MLOps pipelines and strong co-usage patterns, such as between OPA and Gatekeeper. Our taxonomy highlights recurring governance intents. \textbf{Conclusion:} Our findings offer actionable insights for practitioners and tool developers. They highlight concrete usage patterns, emphasize actual PaC usage, and motivate opportunities for improving tool interoperability. This study lays the empirical foundation for future research on PaC practices and their role in ensuring trustworthy, compliant software systems.

cs.SE↗

arXiv2026

Tracing Stereotypes in Pre-trained Transformers: From Biased Neurons to Fairer Models

The advent of transformer-based language models has reshaped how AI systems process and generate text. In software engineering (SE), these models now support diverse activities, accelerating automation and decision-making. Yet, evidence shows that these models can reproduce or amplify social biases, raising fairness concerns. Recent work on neuron editing has shown that internal activations in pre-trained transformers can be traced and modified to alter model behavior. Building on the concept of knowledge neurons, neurons that encode factual information, we hypothesize the existence of biased neurons that capture stereotypical associations within pre-trained transformers. To test this hypothesis, we build a dataset of biased relations, i.e., triplets encoding stereotypes across nine bias types, and adapt neuron attribution strategies to trace and suppress biased neurons in BERT models. We then assess the impact of suppression on SE tasks. Our findings show that biased knowledge is localized within small neuron subsets, and suppressing them substantially reduces bias with minimal performance loss. This demonstrates that bias in transformers can be traced and mitigated at the neuron level, offering an interpretable approach to fairness in SE.

cs.SE↗

arXiv2025

Multi-Agent Framework for Threat Mitigation and Resilience in AI-Based Systems

Machine learning (ML) underpins foundation models in finance, healthcare, and critical infrastructure, making them targets for data poisoning, model extraction, prompt injection, automated jailbreaking, and preference-guided black-box attacks that exploit model comparisons. Larger models can be more vulnerable to introspection-driven jailbreaks and cross-modal manipulation. Traditional cybersecurity lacks ML-specific threat modeling for foundation, multimodal, and RAG systems. Objective: Characterize ML security risks by identifying dominant TTPs, vulnerabilities, and targeted lifecycle stages. Methods: We extract 93 threats from MITRE ATLAS (26), AI Incident Database (12), and literature (55), and analyze 854 GitHub/Python repositories. A multi-agent RAG system (ChatGPT-4o, temp 0.4) mines 300+ articles to build an ontology-driven threat graph linking TTPs, vulnerabilities, and stages. Results: We identify unreported threats including commercial LLM API model stealing, parameter memorization leakage, and preference-guided text-only jailbreaks. Dominant TTPs include MASTERKEY-style jailbreaking, federated poisoning, diffusion backdoors, and preference optimization leakage, mainly impacting pre-training and inference. Graph analysis reveals dense vulnerability clusters in libraries with poor patch propagation. Conclusion: Adaptive, ML-specific security frameworks, combining dependency hygiene, threat intelligence, and monitoring, are essential to mitigate supply-chain and inference risks across the ML lifecycle.

cs.CR↗

arXiv2025

A Survey of Bugs in AI-Generated Code

Developers are widely using AI code-generation models, aiming to increase productivity and efficiency. However, there are also quality concerns regarding the AI-generated code. The generated code is produced by models trained on publicly available code, which are known to contain bugs and quality issues. Those issues can cause trust and maintenance challenges during the development process. Several quality issues associated with AI-generated code have been reported, including bugs and defects. However, these findings are often scattered and lack a systematic summary. A comprehensive review is currently lacking to reveal the types and distribution of these errors, possible remediation strategies, as well as their correlation with the specific models. In this paper, we systematically analyze the existing AI-generated code literature to establish an overall understanding of bugs and defects in generated code, providing a reference for future model improvement and quality assessment. We aim to understand the nature and extent of bugs in AI-generated code, and provide a classification of bug types and patterns present in code generated by different models. We also discuss possible fixes and mitigation strategies adopted to eliminate bugs from the generated code.

cs.SE↗

arXiv2025

The Moral Consistency Pipeline: Continuous Ethical Evaluation for Large Language Models

The rapid advancement and adaptability of Large Language Models (LLMs) highlight the need for moral consistency, the capacity to maintain ethically coherent reasoning across varied contexts. Existing alignment frameworks, structured approaches designed to align model behavior with human ethical and social norms, often rely on static datasets and post-hoc evaluations, offering limited insight into how ethical reasoning may evolve across different contexts or temporal scales. This study presents the Moral Consistency Pipeline (MoCoP), a dataset-free, closed-loop framework for continuously evaluating and interpreting the moral stability of LLMs. MoCoP combines three supporting layers: (i) lexical integrity analysis, (ii) semantic risk estimation, and (iii) reasoning-based judgment modeling within a self-sustaining architecture that autonomously generates, evaluates, and refines ethical scenarios without external supervision. Our empirical results on GPT-4-Turbo and DeepSeek suggest that MoCoP effectively captures longitudinal ethical behavior, revealing a strong inverse relationship between ethical and toxicity dimensions (correlation rET = -0.81, p value less than 0.001) and a near-zero association with response latency (correlation rEL approximately equal to 0). These findings demonstrate that moral coherence and linguistic safety tend to emerge as stable and interpretable characteristics of model behavior rather than short-term fluctuations. Furthermore, by reframing ethical evaluation as a dynamic, model-agnostic form of moral introspection, MoCoP offers a reproducible foundation for scalable, continuous auditing and advances the study of computational morality in autonomous AI systems.

cs.CL↗

arXiv2025

An Empirical Study of Self-Admitted Technical Debt in Machine Learning Software

The emergence of open-source ML libraries such as TensorFlow and Google Auto ML has enabled developers to harness state-of-the-art ML algorithms with minimal overhead. However, during this accelerated ML development process, said developers may often make sub-optimal design and implementation decisions, leading to the introduction of technical debt that, if not addressed promptly, can have a significant impact on the quality of the ML-based software. Developers frequently acknowledge these sub-optimal design and development choices through code comments during software development. These comments, which often highlight areas requiring additional work or refinement in the future, are known as self-admitted technical debt (SATD). This paper aims to investigate SATD in ML code by analyzing 318 open-source ML projects across five domains, along with 318 non-ML projects. We detected SATD in source code comments throughout the different project snapshots, conducted a manual analysis of the identified SATD sample to comprehend the nature of technical debt in the ML code, and performed a survival analysis of the SATD to understand the evolution of such debts. We observed: i) Machine learning projects have a median percentage of SATD that is twice the median percentage of SATD in non-machine learning projects. ii) ML pipeline components for data preprocessing and model generation logic are more susceptible to debt than model validation and deployment components. iii) SATDs appear in ML projects earlier in the development process compared to non-ML projects. iv) Long-lasting SATDs are typically introduced during extensive code changes that span multiple files exhibiting low complexity.

cs.SE↗

arXiv2025

Think Fast: Real-Time IoT Intrusion Reasoning Using IDS and LLMs at the Edge Gateway

As the number of connected IoT devices continues to grow, securing these systems against cyber threats remains a major challenge, especially in environments with limited computational and energy resources. This paper presents an edge-centric Intrusion Detection System (IDS) framework that integrates lightweight machine learning (ML) based IDS models with pre-trained large language models (LLMs) to improve detection accuracy, semantic interpretability, and operational efficiency at the network edge. The system evaluates six ML-based IDS models: Decision Tree (DT), K-Nearest Neighbors (KNN), Random Forest (RF), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM model on low-power edge gateways, achieving accuracy up to 98 percent under real-world cyberattacks. For anomaly detection, the system transmits a compact and secure telemetry snapshot (for example, CPU usage, memory usage, latency, and energy consumption) via low-bandwidth API calls to LLMs including GPT-4-turbo, DeepSeek V2, and LLaMA 3.5. These models use zero-shot, few-shot, and chain-of-thought reasoning to produce human-readable threat analyses and actionable mitigation recommendations. Evaluations across diverse attacks such as DoS, DDoS, brute force, and port scanning show that the system enhances interpretability while maintaining low latency (<1.5 s), minimal bandwidth usage (<1.2 kB per prompt), and energy efficiency (<75 J), demonstrating its practicality and scalability as an IDS solution for edge gateways.

cs.CR↗

arXiv2025

One Size Does Not Fit All: Architecture-Aware Adaptive Batch Scheduling with DEBA

Adaptive batch size methods aim to accelerate neural network training, but existing approaches apply identical adaptation strategies across all architectures, assuming a one-size-fits-all solution. We introduce DEBA (Dynamic Efficient Batch Adaptation), an adaptive batch scheduler that monitors gradient variance, gradient norm variation and loss variation to guide batch size adaptations. Through systematic evaluation across six architectures (ResNet-18/50, DenseNet-121, EfficientNet-B0, MobileNet-V3, ViT-B16) on CIFAR-10 and CIFAR-100, with five random seeds per configuration, we demonstrate that the architecture fundamentally determines adaptation efficacy. Our findings reveal that: (1) lightweight and medium-depth architectures (MobileNet-V3, DenseNet-121, EfficientNet-B0) achieve a 45-62% training speedup with simultaneous accuracy improvements of 1-7%; (2) shallow residual networks (ResNet-18) show consistent gains of +2.4 - 4.0% in accuracy, 36 - 43% in speedup, while deep residual networks (ResNet-50) exhibit high variance and occasional degradation; (3) already-stable architectures (ViT-B16) show minimal speedup (6%) despite maintaining accuracy, indicating that adaptation benefits vary with baseline optimization characteristics. We introduce a baseline characterization framework using gradient stability metrics (stability score, gradient norm variation) that predicts which architectures will benefit from adaptive scheduling. Our ablation studies reveal critical design choices often overlooked in prior work: sliding window statistics (vs. full history) and sufficient cooldown periods (5+ epochs) between adaptations are essential for success. This work challenges the prevailing assumption that adaptive methods generalize across architectures and provides the first systematic evidence that batch size adaptation requires an architecture-aware design.

cs.LG↗

arXiv2025

TaskEval: Assessing Difficulty of Code Generation Tasks for Large Language Models

Large Language Models (LLMs) excel in code-related tasks like code generation, but benchmark evaluations often overlook task characteristics, such as difficulty. Moreover, benchmarks are usually built using tasks described with a single prompt, despite the formulation of prompts having a profound impact on the outcome. This paper introduces a generalist approach, TaskEval, a framework using diverse prompts and Item Response Theory (IRT) to efficiently assess LLMs' capabilities and benchmark task characteristics, improving the understanding of their performance. Using two code generation benchmarks, \textit{HumanEval}+ and \textit{ClassEval}, as well as 8 code generation LLMs, we show that \textit{TaskEval} is capable of characterising the properties of tasks. Using topic analysis, we identify and analyse the tasks of 17 and 21 topics within the benchmarks. We also cross-analyse tasks' characteristics with programming constructs (e.g., variable assignment, conditions, etc.) used by LLMs, emphasising some patterns with tasks' difficulty. Finally, we conduct a comparison between the difficulty assessment of tasks by human annotators and LLMs. Orthogonal to current benchmarking evaluation efforts, \textit{TaskEval} can assist researchers and practitioners in fostering better assessments of LLMs. The tasks' characteristics can be used to identify shortcomings within existing benchmarks or improve the evaluation of LLMs.

cs.SE↗

arXiv2025

Refactoring with LLMs: Bridging Human Expertise and Machine Understanding

Code refactoring is a fundamental software engineering practice aimed at improving code quality and maintainability. Despite its importance, developers often neglect refactoring due to the significant time, effort, and resources it requires, as well as the lack of immediate functional rewards. Although several automated refactoring tools have been proposed, they remain limited in supporting a broad spectrum of refactoring types. In this study, we explore whether instruction strategies inspired by human best-practice guidelines can enhance the ability of Large Language Models (LLMs) to perform diverse refactoring tasks automatically. Leveraging the instruction-following and code comprehension capabilities of state-of-the-art LLMs (e.g., GPT-mini and DeepSeek-V3), we draw on Martin Fowler's refactoring guidelines to design multiple instruction strategies that encode motivations, procedural steps, and transformation objectives for 61 well-known refactoring types. We evaluate these strategies on benchmark examples and real-world code snippets from GitHub projects. Our results show that instruction designs grounded in Fowler's guidelines enable LLMs to successfully perform all benchmark refactoring types and preserve program semantics in real-world settings, an essential criterion for effective refactoring. Moreover, while descriptive instructions are more interpretable to humans, our results show that rule-based instructions often lead to better performance in specific scenarios. Interestingly, allowing models to focus on the overall goal of refactoring, rather than prescribing a fixed transformation type, can yield even greater improvements in code quality.

cs.SE↗

arXiv2025

Logging Requirement for Continuous Auditing of Responsible Machine Learning-based Applications

Machine learning (ML) is increasingly applied across industries to automate decision-making, but concerns about ethical and legal compliance remain due to limited transparency, fairness, and accountability. Monitoring through logging a long-standing practice in traditional software offers a potential means for auditing ML applications, as logs provide traceable records of system behavior useful for debugging, performance analysis, and continuous auditing. systematically auditing models for compliance or accountability. The findings underscore the need for enhanced logging practices and tooling that systematically integrate responsible AI metrics. Such practices would support the development of auditable, transparent, and ethically responsible ML systems, aligning with growing regulatory requirements and societal expectations. By highlighting specific deficiencies and opportunities, this work provides actionable guidance for both practitioners and tool developers seeking to strengthen the accountability and trustworthiness of ML applications.

cs.SE↗

arXiv2025

Towards Understanding the Impact of Data Bugs on Deep Learning Models in Software Engineering

Deep learning (DL) techniques have achieved significant success in various software engineering tasks (e.g., code completion by Copilot). However, DL systems are prone to bugs from many sources, including training data. Existing literature suggests that bugs in training data are highly prevalent, but little research has focused on understanding their impacts on the models used in software engineering tasks. In this paper, we address this research gap through a comprehensive empirical investigation focused on three types of data prevalent in software engineering tasks: code-based, text-based, and metric-based. Using state-of-the-art baselines, we compare the models trained on clean datasets with those trained on datasets with quality issues and without proper preprocessing. By analysing the gradients, weights, and biases from neural networks under training, we identify the symptoms of data quality and preprocessing issues. Our analysis reveals that quality issues in code data cause biased learning and gradient instability, whereas problems in text data lead to overfitting and poor generalisation of models. On the other hand, quality issues in metric data result in exploding gradients and model overfitting, and inadequate preprocessing exacerbates these effects across all three data types. Finally, we demonstrate the validity and generalizability of our findings using six new datasets. Our research provides a better understanding of the impact and symptoms of data bugs in software engineering datasets. Practitioners and researchers can leverage these findings to develop better monitoring systems and data-cleaning methods to help detect and resolve data bugs in deep learning systems.

cs.SE↗

arXiv2025

FairFLRep: Fairness aware fault localization and repair of Deep Neural Networks

Deep neural networks (DNNs) are being utilized in various aspects of our daily lives, including high-stakes decision-making applications that impact individuals. However, these systems reflect and amplify bias from the data used during training and testing, potentially resulting in biased behavior and inaccurate decisions. For instance, having different misclassification rates between white and black sub-populations. However, effectively and efficiently identifying and correcting biased behavior in DNNs is a challenge. This paper introduces FairFLRep, an automated fairness-aware fault localization and repair technique that identifies and corrects potentially bias-inducing neurons in DNN classifiers. FairFLRep focuses on adjusting neuron weights associated with sensitive attributes, such as race or gender, that contribute to unfair decisions. By analyzing the input-output relationships within the network, FairFLRep corrects neurons responsible for disparities in predictive quality parity. We evaluate FairFLRep on four image classification datasets using two DNN classifiers, and four tabular datasets with a DNN model. The results show that FairFLRep consistently outperforms existing methods in improving fairness while preserving accuracy. An ablation study confirms the importance of considering fairness during both fault localization and repair stages. Our findings also show that FairFLRep is more efficient than the baseline approaches in repairing the network.

cs.LG↗