arXiv subjects

Vishnu Asutosh Dasu

Publications and source records attributed to Vishnu Asutosh Dasu.

11 recordsLinked to original sources

arXiv2026

Refusing the Impossible: A Taxonomy and Benchmark for Code Hallucination in Large Language Models

Large language models (LLMs) often produce code that looks plausible but is not grounded in reality. The code may import packages that do not exist or claim to implement algorithms that violate proven theorems, while still compiling and running. We study \emph{code hallucination} as \emph{ungrounded generation} and separate it from ordinary \emph{code error} (bugs in otherwise grounded programs). We propose a taxonomy with three dimensions: \textbf{groundedness} (absolute violations of universal truths vs.\ relative fabrications of contingent or ecosystem-specific facts), \textbf{manifestation level} (syntactic, semantic, or factual), and \textbf{behavior} (from confident fabrication to degenerate output), organized into a severity ordering. We build an \textbf{adversarial} suite of deliberately unsatisfiable tasks where the correct response is to refuse and categorize the responses under our taxonomy. The suite contains \textbf{270 prompts} across six languages and 24 subcategories, paired with \textbf{91 matched solvable controls}, and responses are judged by a two-tier protocol validated against human labels (82\% agreement, $κ{=}0.73$). Across twelve open-weight code and reasoning models (4{,}332 judged responses), models produce ungrounded code on about 60\% of unsatisfiable prompts and refuse only 27\%, while wrongly refusing 0\% of the solvable controls.

cs.SE↗

arXiv2026

Robustness of Vision Language Models Against Split-Image Harmful Input Attacks

Vision-Language Models (VLMs) are now a core part of modern AI. Recent work proposed several visual jailbreak attacks using single/ holistic images. However, contemporary VLMs demonstrate strong robustness against such attacks due to extensive safety alignment through preference optimization, e.g., reinforcement learning from human feedback (RLHF). In this work, we identify a new vulnerability: while VLM pretraining and instruction tuning generalize well to split-image inputs, safety alignment is typically performed only on holistic images and does not account for harmful semantics distributed across multiple image fragments. Consequently, VLMs often fail to detect and reject harmful split-image inputs, in which unsafe cues emerge only upon combining images. We introduce novel split-image visual jailbreak attacks (\textbf{SIVA}) that exploit this misalignment. Unlike prior optimization-based attacks, which exhibit poor black-box transferability due to architectural and prior mismatches across models, our attacks evolve in progressive phases from naive splitting to an adaptive white-box attack, culminating in a black-box transfer attack. Our strongest strategy leverages a novel adversarial knowledge distillation \textbf{(Adv-KD)} algorithm to substantially improve cross-model transferability. Evaluations on four state-of-the-art modern VLMs and three jailbreak datasets demonstrate that our strongest attack achieves up to 44% higher transfer success than existing baselines. Lastly, we propose efficient ways to address this critical vulnerability in the current VLM safety alignment.

cs.CV↗

arXiv2026

Heimdall: Formally Verified Automated Migration of Legacy eBPF Programs to Rust

Extended Berkeley Packet Filter (eBPF) programs are kernel extensions used for networking, observability, and security enforcement in the Linux kernel. The in-kernel eBPF verifier checks low-level memory safety and termination on eBPF programs, but it does not enforce many higher-level source-level properties, such as initialization discipline, schema consistency, or error handling. We document nine classes of source-level bugs that compile, pass the kernel verifier, and can silently corrupt data, leak kernel memory to userspace, or yield incorrect enforcement outcomes. To harden such verifier-accepted buggy programs and support safe migration, we present Heimdall, an automated pipeline that uses large language models to translate legacy libbpf C programs to Aya Rust. Heimdall iteratively repairs compilation and kernel-verifier failures, rejects unsafe escape hatches in Rust-Aya with a static analysis safety engine, and proves per-program equivalence to the original via symbolic execution and Z3-based equivalence checking. Across 115 eBPF programs, Heimdall generates 109 formally proven-equivalent translations (94.8%). In the process, Heimdall identifies nine bugs in real-world eBPF programs and fixes all of them, two of which leak randomized kernel addresses to userspace and break KASLR. Eight of the nine have also been acknowledged and fixed upstream by the developers. Heimdall is the first system to automate memory-safe-language migration of production eBPF programs with per-program formal guarantees that the migration preserves observable behavior.

cs.CR↗

arXiv2026

Gradient-Free Privacy Leakage in Federated Language Models through Selective Weight Tampering

Federated learning (FL) has become a key component in various language modeling applications such as machine translation, next-word prediction, and medical record analysis. These applications are trained on datasets from many FL participants that often include privacy-sensitive data, such as healthcare records, phone/credit card numbers, login credentials, etc. Although FL enables computation without necessitating clients to share their raw data, existing works show that privacy leakage is still probable in federated language models. In this paper, we present two novel findings on the leakage of privacy-sensitive user data from federated large language models without requiring access to gradients. Firstly, we make a key observation that model snapshots from the intermediate rounds in FL can cause greater privacy leakage than the final trained model. Secondly, we identify that a malicious FL participant can aggravate the leakage by tampering with the model's selective weights that are responsible for memorizing the sensitive training data of some other clients, even without any cooperation from the server. Our best-performing method increases the membership inference recall by 29% and achieves up to 71% private data reconstruction, evidently outperforming existing attacks that consider much stronger adversary capabilities. Lastly, we recommend a balanced suite of techniques for an FL client to defend against such privacy risk.

cs.CR↗

arXiv2026

Impact of Data Duplication on Deep Neural Network-Based Image Classifiers: Robust vs. Standard Models

The accuracy and robustness of machine learning models against adversarial attacks are significantly influenced by factors such as training data quality, model architecture, the training process, and the deployment environment. In recent years, duplicated data in training sets, especially in language models, has attracted considerable attention. It has been shown that deduplication enhances both training performance and model accuracy in language models. While the importance of data quality in training image classifier Deep Neural Networks (DNNs) is widely recognized, the impact of duplicated images in the training set on model generalization and performance has received little attention. In this paper, we address this gap and provide a comprehensive study on the effect of duplicates in image classification. Our analysis indicates that the presence of duplicated images in the training set not only negatively affects the efficiency of model training but also may result in lower accuracy of the image classifier. This negative impact of duplication on accuracy is particularly evident when duplicated data is non-uniform across classes or when duplication, whether uniform or non-uniform, occurs in the training set of an adversarially trained model. Even when duplicated samples are selected in a uniform way, increasing the amount of duplication does not lead to a significant improvement in accuracy.

cs.LG↗

arXiv2026

iResolveX: Multi-Layered Indirect Call Resolution via Static Reasoning and Learning-Augmented Refinement

Indirect call resolution remains a key challenge in reverse engineering and control-flow graph recovery, especially for stripped or optimized binaries. Static analysis is sound but often over-approximates, producing many false positives, whereas machine-learning approaches can improve precision but may sacrifice completeness and generalization. We present iResolveX, a hybrid multi-layered framework that combines conservative static analysis with learning-based refinement. The first layer applies a conservative value-set analysis (BPA) to ensure high recall. The second layer adds a learning-based soft-signature scorer (iScoreGen) and selective inter-procedural backward analysis with memory inspection (iScoreRefine) to reduce false positives. The final output, p-IndirectCFG, annotates indirect edges with confidence scores, enabling downstream analyses to choose appropriate precision--recall trade-offs. Across SPEC CPU2006 and real-world binaries, iScoreGen reduces predicted targets by 19.2% on average while maintaining BPA-level recall (98.2%). Combined with iScoreRefine, the total reduction reaches 44.3% over BPA with 97.8% recall (a 0.4% drop). iResolveX supports both conservative, recall-preserving and F1-optimized configurations and outperforms state-of-the-art systems.

cs.SE↗

arXiv2025

Attention Pruning: Automated Fairness Repair of Language Models via Surrogate Simulated Annealing

This paper explores pruning attention heads as a post-processing bias mitigation method for large language models (LLMs). Modern AI systems such as LLMs are expanding into sensitive social contexts where fairness concerns become especially crucial. Since LLMs develop decision-making patterns by training on massive datasets of human-generated content, they naturally encode and perpetuate societal biases. While modifying training datasets and algorithms is expensive and requires significant resources; post-processing techniques-such as selectively deactivating neurons and attention heads in pre-trained LLMs-can provide feasible and effective approaches to improve fairness. However, identifying the optimal subset of parameters to prune presents a combinatorial challenge within LLMs' immense parameter space, requiring solutions that efficiently balance competing objectives across the frontiers of model fairness and utility. To address the computational challenges, we explore a search-based program repair approach via randomized simulated annealing. Given the prohibitive evaluation costs in billion-parameter LLMs, we develop surrogate deep neural networks that efficiently model the relationship between attention head states (active/inactive) and their corresponding fairness/utility metrics. This allows us to perform optimization over the surrogate models and efficiently identify optimal subsets of attention heads for selective pruning rather than directly searching through the LLM parameter space. This paper introduces Attention Pruning, a fairness-aware surrogate simulated annealing approach to prune attention heads in LLMs that disproportionately contribute to bias while minimally impacting overall model utility. Our experiments show that Attention Pruning achieves up to $40\%$ reduction in gender bias and outperforms the state-of-the-art bias mitigation strategies.

cs.AI↗

arXiv2025

Chain-of-Thought Driven Adversarial Scenario Extrapolation for Robust Language Models

Large Language Models (LLMs) exhibit impressive capabilities, but remain susceptible to a growing spectrum of safety risks, including jailbreaks, toxic content, hallucinations, and bias. Existing defenses often address only a single threat type or resort to rigid outright rejection, sacrificing user experience and failing to generalize across diverse and novel attacks. This paper introduces Adversarial Scenario Extrapolation (ASE), a novel inference-time computation framework that leverages Chain-of-Thought (CoT) reasoning to simultaneously enhance LLM robustness and seamlessness. ASE guides the LLM through a self-generative process of contemplating potential adversarial scenarios and formulating defensive strategies before generating a response to the user query. Comprehensive evaluation on four adversarial benchmarks with four latest LLMs shows that ASE achieves near-zero jailbreak attack success rates and minimal toxicity, while slashing outright rejections to <4%. ASE outperforms six state-of-the-art defenses in robustness-seamlessness trade-offs, with 92-99% accuracy on adversarial Q&A and 4-10x lower bias scores. By transforming adversarial perception into an intrinsic cognitive process, ASE sets a new paradigm for secure and natural human-AI interaction.

cs.CL↗

arXiv2025

Improving Noise Efficiency in Privacy-preserving Dataset Distillation

Modern machine learning models heavily rely on large datasets that often include sensitive and private information, raising serious privacy concerns. Differentially private (DP) data generation offers a solution by creating synthetic datasets that limit the leakage of private information within a predefined privacy budget; however, it requires a substantial amount of data to achieve performance comparable to models trained on the original data. To mitigate the significant expense incurred with synthetic data generation, Dataset Distillation (DD) stands out for its remarkable training and storage efficiency. This efficiency is particularly advantageous when integrated with DP mechanisms, curating compact yet informative synthetic datasets without compromising privacy. However, current state-of-the-art private DD methods suffer from a synchronized sampling-optimization process and the dependency on noisy training signals from randomly initialized networks. This results in the inefficient utilization of private information due to the addition of excessive noise. To address these issues, we introduce a novel framework that decouples sampling from optimization for better convergence and improves signal quality by mitigating the impact of DP noise through matching in an informative subspace. On CIFAR-10, our method achieves a \textbf{10.0\%} improvement with 50 images per class and \textbf{8.3\%} increase with just \textbf{one-fifth} the distilled set size of previous state-of-the-art methods, demonstrating significant potential to advance privacy-preserving DD.

cs.CV↗

arXiv2024

Privacy-Preserving Data Deduplication for Enhancing Federated Learning of Language Models (Extended Version)

Deduplication is a vital preprocessing step that enhances machine learning model performance and saves training time and energy. However, enhancing federated learning through deduplication poses challenges, especially regarding scalability and potential privacy violations if deduplication involves sharing all clients' data. In this paper, we address the problem of deduplication in a federated setup by introducing a pioneering protocol, Efficient Privacy-Preserving Multi-Party Deduplication (EP-MPD). It efficiently removes duplicates from multiple clients' datasets without compromising data privacy. EP-MPD is constructed in a modular fashion, utilizing two novel variants of the Private Set Intersection protocol. Our extensive experiments demonstrate the significant benefits of deduplication in federated learning of large language models. For instance, we observe up to 19.62\% improvement in perplexity and up to 27.95\% reduction in running time while varying the duplication level between 10\% and 30\%. EP-MPD effectively balances privacy and performance in federated learning, making it a valuable solution for large-scale applications.

cs.CR↗

arXiv2024

NeuFair: Neural Network Fairness Repair with Dropout

This paper investigates neuron dropout as a post-processing bias mitigation for deep neural networks (DNNs). Neural-driven software solutions are increasingly applied in socially critical domains with significant fairness implications. While neural networks are exceptionally good at finding statistical patterns from data, they may encode and amplify existing biases from the historical data. Existing bias mitigation algorithms often require modifying the input dataset or the learning algorithms. We posit that the prevalent dropout methods that prevent over-fitting during training by randomly dropping neurons may be an effective and less intrusive approach to improve the fairness of pre-trained DNNs. However, finding the ideal set of neurons to drop is a combinatorial problem. We propose NeuFair, a family of post-processing randomized algorithms that mitigate unfairness in pre-trained DNNs via dropouts during inference after training. Our randomized search is guided by an objective to minimize discrimination while maintaining the model's utility. We show that our design of randomized algorithms is effective and efficient in improving fairness (up to 69%) with minimal or no model performance degradation. We provide intuitive explanations of these phenomena and carefully examine the influence of various hyperparameters of search algorithms on the results. Finally, we empirically and conceptually compare NeuFair to different state-of-the-art bias mitigators.

cs.LG↗