Searcharxiv⌕ Search

arXiv subjects

Nael Abu-Ghazaleh

Publications and source records attributed to Nael Abu-Ghazaleh.

At least 19 recordsLinked to original sources
arXiv2026

PerSpectron: Detecting Invariant Footprints of Microarchitectural Attacks with Perceptron

Detecting microarchitectural attacks is critical given their proliferation in recent years. Many of these attacks exhibit intrinsic behaviors essential to the nature of their operation, such as creating contention or misspeculation. This study systematically investigates the microarchitectural footprints of hardware-based attacks and shows how they can be detected and classified using an efficient hardware predictor. We present a methodology to use correlated microarchitectural statistics to design a hardware-based neural predictor capable of detecting and classifying microarchitectural attacks before data is leaked. Once a potential attack is detected, it can be proactively mitigated by triggering appropriate countermeasures. Our hardware-based detector, PerSpectron, uses perceptron learning to identify and classify attacks. Perceptron-based prediction has been successfully used in branch prediction and other hardware-based applications. PerSpectron has minimal performance overhead. The statistics being monitored have similar overhead to already existing performance monitoring counters. Additionally, PerSpectron operates outside the processor's critical paths, offering security without added computation delay. Our system achieves a usable detection rate for detecting attacks such as SpectreV1, SpectreV2, SpectreRSB, Meltdown, breakingKSLR, Flush+Flush, Flush+Reload, Prime+Probe as well as cache-attack calibration programs. We also believe that the large number of diverse microarchitectural features offers both evasion resilience and interpretability---features not present in previous hardware security detectors. We detect these attacks early enough to avoid any data leakage, unlike previous work that triggers countermeasures only after data has been exposed.

cs.CR↗
arXiv2026

Visual Grounding Safety in Vision-Language Models

Vision-language models (VLMs) are increasingly trained to generate structured outputs like points and bounding boxes that downstream interfaces, agents, and robots can act on, yet safety alignment of this output channel has not been systematically analyzed. We study visual grounding safety by repurposing three safety benchmarks spanning direct harm (VLSU), social bias (BBQ-V), and situational safety (Asimov-2.0) into 15,401 matched pairs of harmful requests that differ only in the requested output: a free-text answer (VQA) or a grounding (point or bounding box). Across five VLMs, models that refuse a harmful request posed as a question often comply when the same request asks for a grounding: averaged over models, grounding refusal trails VQA refusal by 31-59 percentage points, depending on the domain, and safety system prompts do not close this gap. We propose a fine-tuning approach that combines grounding-form refusals with capability grounding data and self-distilled benign data to counter over-refusal. For Qwen3-VL-8B and VisionReasoner-7B, it improves grounding refusal by 77-95 percentage points on VLSU and BBQ-V and by 64-85 points on the held-out Asimov-2.0 domain, while also improving VQA refusal, preserving grounding capability, and keeping over-refusal limited. Representation analysis shows that fine-tuning moves harmful requests toward each model's refusal direction, most strongly for grounding, while leaving benign requests near the harmless reference.

cs.AI↗
arXiv2026

Microflow: Microarchitectural Causal Observability for Deep Cross-Layer Analysis and Optimization

Modern computer architecture relies heavily on simulation to identify bottlenecks and evaluate optimizations. However, existing microarchitectural performance analysis methods are fundamentally limited by an instruction-centric paradigm that captures only downstream symptoms while leaving the true microarchitectural root cause opaque. Because modern processors are governed by complex interactions across non-instruction entities like prefetchers, replacement policies, and shared queue occupancies, instruction-centric frameworks miss the mechanisms that dictate performance. To eliminate this blind spot, we present Microflow, a framework that achieves causal observability in microarchitectural simulation. To address this, we introduce the Microflow intermediate representation (MFIR), which models execution through microarchitecture-tailored core abstractions such as flows, resource tenancies, and causal edges. By compiling simulation runs into a relational causal database, Microflow decouples tracing from analytical processing. This transforms complex diagnostics into expressive queries, enabling architects to trace performance symptoms directly to hardware root causes without developing bespoke analysis scripts or running costly re-simulations for every new question. We demonstrate that Microflow solves pathologies opaque to conventional tools. Across CVP-1 benchmarks, Microflow decomposes a 22% prefetcher oracle headroom by attributing 61.4% of stall mass to specific hardware prefetcher decisions, yielding a 2.31% average speedup (peaking at 25.11%). Furthermore, it exposes the hidden pipeline-blocking residue of wrong-path execution with high portability and precision across simulators.

cs.AR↗
arXiv2026

VLMs Need Words: Vision Language Models Ignore Visual Detail In Favor of Semantic Anchors

Vision-language models (VLMs) have achieved impressive performance across a wide range of multimodal tasks. However, they often fail on tasks that require fine-grained visual perception, even when the required information is still present in their internal representations. Prior work has attributed this ``hidden-in-plain-sight'' gap to the language model, but the cause remains unexplained. In this work, we demonstrate that this gap arises from the language model's lack of semantic labels for fine-grained visual details: when visual entities can be mapped to known concepts, VLMs bypass visual comparison and reason through language; when they cannot, VLMs resort to brittle and hallucinated descriptions. We verify this across semantic correspondence, synthetic shape matching, and face matching, and find that VLMs perform much better when the relevant entities are nameable than when they are unnamable. Mechanistically, Logit Lens analysis confirms that VLMs explicitly recover semantic labels for nameable entities and surface more unique tokens compared to unnameable entities. Furthermore, we show that this limitation can be addressed: teaching completely arbitrary names for unknown entities improves performance. More importantly, task-specific finetuning yields even stronger generalization without relying on language priors, i.e., through real visual perception. Our findings suggest that current VLM failures on visual tasks reflect a learned shortcut rather than a fundamental limitation of multimodal reasoning. Code and datasets are available at https://github.com/Patchwork53/VLMs-Need-Words-COLM2026.

cs.CV↗
arXiv2026

Loss Landscape Poisoning: Targeted Extraction of Unseen Training Data from LLMs

Large Language Models are increasingly trained on proprietary or sensitive data, from private healthcare and financial records to user conversations containing secrets. Ensuring the privacy of such data against extraction attacks has become a central concern. In this paper, we ask whether an attacker who can poison a portion of the training data can facilitate the leakage of a separate target record they have no access to. We answer in the affirmative and show that such leakage can be induced by a poisoning mechanism that reshapes the model's local loss landscape around the target completion. Our key insight is that poisoning to create a sharp loss minimum at the target, surrounded by elevated loss on nearby alternatives, forces the model to memorize the target as the unique low-loss solution in its neighborhood. The attack requires no architectural changes, and generalizes across centralized and federated learning settings. We demonstrate that the attack amplifies privacy leakage across language (up to 100% successful extraction), and vision-language models (up 90% successful extraction). We show that the attack is thwarted when the model is trained to be differentially private. However, we introduce a new attack that directly probes the loss landscape bypassing even differential privacy defenses.

cs.CR↗
arXiv2026

Modeling Hierarchical Thinking in Large Reasoning Models

Large Reasoning Models (LRMs) solve complex tasks by generating long Chain-of-Thought (CoT) sequences; however, the emergent dynamics governing reasoning trajectories are not well understood and can lead to inconsistencies and reasoning pathologies. In this work, we propose to approximate LRM's emerging hierarchical reasoning dynamics as a trajectory within a Finite State Machine (FSM) transitioning among six abstract cognitive states. We demonstrate that these states and transitions can be captured in the latent state of the model. We believe that this representation can have different applications in the interpretability and optimization of LRM models. For example, by analyzing the topology of these transitions, we identify statistical shifts in reasoning strategies that help identify effective reasoning chains from those that fail. To illustrate these potential advantages, we propose Q-Value guided steering, a training-free inference-time control method that treats reasoning as a planning problem. We estimate the long-horizon utility of state transitions and apply sparse, orthogonal activation steering at sentence boundaries to align the CoT generation with optimal reasoning policies. Experiments across four benchmarks (AIME25, MATH-500, GSM8k, and GPQA Diamond) using three state-of-the-art open reasoning models demonstrate that Q-Value steering policy achieves significant performance gains with "surgical" efficiency, often requiring 25 times fewer interventions than greedy and weighted baselines, which suggests that reasoning can be effectively controlled by guiding high-level cognitive dynamics rather than micro-managing token generation. Code is available at: https://github.com/shahariar-shibli/CoT-FSM.

cs.AI↗
arXiv2026

AttenMIA: LLM Membership Inference Attack through Attention Signals

Large Language Models (LLMs) are increasingly deployed to enable or improve a multitude of real-world applications. Given the large size of their training data sets, their tendency to memorize training data raises serious privacy and intellectual property concerns. A key threat is the membership inference attack (MIA), which aims to determine whether a given sample was included in the model's training set. Existing MIAs for LLMs rely primarily on output confidence scores or embedding-based features, but these signals are often brittle, leading to limited attack success. We introduce AttenMIA, a new MIA framework that exploits self-attention patterns inside the transformer model to infer membership. Attention controls the information flow within the transformer, exposing different patterns for memorization that can be used to identify members of the dataset. Our method uses information from attention heads across layers and combines them with perturbation-based divergence metrics to train an effective MIA classifier. Using extensive experiments on open-source models including LLaMA-2, Pythia, and Opt models, we show that attention-based features consistently outperform baselines, particularly under the important low-false-positive metric (e.g., achieving up to 0.996 ROC AUC & 87.9% TPR@1%FPR on the WikiMIA-32 benchmark with Llama2-13b). We show that attention signals generalize across datasets and architectures, and provide a layer- and head-level analysis of where membership leakage is most pronounced. We also show that using AttenMIA to replace other membership inference attacks in a data extraction framework results in training data extraction attacks that outperform the state of the art. Our findings reveal that attention mechanisms, originally introduced to enhance interpretability, can inadvertently amplify privacy risks in LLMs, underscoring the need for new defenses.

cs.LG↗
arXiv2025

Cross-Modal Safety Alignment: Is textual unlearning all you need?

Recent studies reveal that integrating new modalities into Large Language Models (LLMs), such as Vision-Language Models (VLMs), creates a new attack surface that bypasses existing safety training techniques like Supervised Fine-tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF). While further SFT and RLHF-based safety training can be conducted in multi-modal settings, collecting multi-modal training datasets poses a significant challenge. Inspired by the structural design of recent multi-modal models, where, regardless of the combination of input modalities, all inputs are ultimately fused into the language space, we aim to explore whether unlearning solely in the textual domain can be effective for cross-modality safety alignment. Our evaluation across six datasets empirically demonstrates the transferability -- textual unlearning in VLMs significantly reduces the Attack Success Rate (ASR) to less than 8\% and in some cases, even as low as nearly 2\% for both text-based and vision-text-based attacks, alongside preserving the utility. Moreover, our experiments show that unlearning with a multi-modal dataset offers no potential benefits but incurs significantly increased computational demands, possibly up to 6 times higher.

cs.CL↗
arXiv2025

Just Do It!? Computer-Use Agents Exhibit Blind Goal-Directedness

Computer-Use Agents (CUAs) are an increasingly deployed class of agents that take actions on GUIs to accomplish user goals. In this paper, we show that CUAs consistently exhibit Blind Goal-Directedness (BGD): a bias to pursue goals regardless of feasibility, safety, reliability, or context. We characterize three prevalent patterns of BGD: (i) lack of contextual reasoning, (ii) assumptions and decisions under ambiguity, and (iii) contradictory or infeasible goals. We develop BLIND-ACT, a benchmark of 90 tasks capturing these three patterns. Built on OSWorld, BLIND-ACT provides realistic environments and employs LLM-based judges to evaluate agent behavior, achieving 93.75% agreement with human annotations. We use BLIND-ACT to evaluate nine frontier models, including Claude Sonnet and Opus 4, Computer-Use-Preview, and GPT-5, observing high average BGD rates (80.8%) across them. We show that BGD exposes subtle risks that arise even when inputs are not directly harmful. While prompting-based interventions lower BGD levels, substantial risk persists, highlighting the need for stronger training- or inference-time interventions. Qualitative analysis reveals observed failure modes: execution-first bias (focusing on how to act over whether to act), thought-action disconnect (execution diverging from reasoning), and request-primacy (justifying actions due to user request). Identifying BGD and introducing BLIND-ACT establishes a foundation for future research on studying and mitigating this fundamental risk and ensuring safe CUA deployment.

cs.AI↗
arXiv2025

Evil Vizier: Vulnerabilities of LLM-Integrated XR Systems

Extended reality (XR) applications increasingly integrate Large Language Models (LLMs) to enhance user experience, scene understanding, and even generate executable XR content, and are often called "AI glasses". Despite these potential benefits, the integrated XR-LLM pipeline makes XR applications vulnerable to new forms of attacks. In this paper, we analyze LLM-Integated XR systems in the literature and in practice and categorize them along different dimensions from a systems perspective. Building on this categorization, we identify a common threat model and demonstrate a series of proof-of-concept attacks on multiple XR platforms that employ various LLM models (Meta Quest 3, Meta Ray-Ban, Android, and Microsoft HoloLens 2 running Llama and GPT models). Although these platforms each implement LLM integration differently, they share vulnerabilities where an attacker can modify the public context surrounding a legitimate LLM query, resulting in erroneous visual or auditory feedback to users, thus compromising their safety or privacy, sowing confusion, or other harmful effects. To defend against these threats, we discuss mitigation strategies and best practices for developers, including an initial defense prototype, and call on the community to develop new protection mechanisms to mitigate these risks.

cs.CR↗
arXiv2025

ShadowScope: GPU Monitoring and Validation via Composable Side Channel Signals

As modern systems increasingly rely on GPUs for computationally intensive tasks such as machine learning acceleration, ensuring the integrity of GPU computation has become critically important. Recent studies have shown that GPU kernels are vulnerable to both traditional memory safety issues (e.g., buffer overflow attacks) and emerging microarchitectural threats (e.g., Rowhammer attacks), many of which manifest as anomalous execution behaviors observable through side-channel signals. However, existing golden model based validation approaches that rely on such signals are fragile, highly sensitive to interference, and do not scale well across GPU workloads with diverse scheduling behaviors. To address these challenges, we propose ShadowScope, a monitoring and validation framework that leverages a composable golden model. Instead of building a single monolithic reference, ShadowScope decomposes trusted kernel execution into modular, repeatable functions that encode key behavioral features. This composable design captures execution patterns at finer granularity, enabling robust validation that is resilient to noise, workload variation, and interference across GPU workloads. To further reduce reliance on noisy software-only monitoring, we introduce ShadowScope+, a hardware-assisted validation mechanism that integrates lightweight on-chip checks into the GPU pipeline. ShadowScope+ achieves high validation accuracy with an average runtime overhead of just 4.6%, while incurring minimal hardware and design complexity. Together, these contributions demonstrate that side-channel observability can be systematically repurposed into a practical defense for GPU kernel integrity.

cs.CR↗
arXiv2025

Poison Once, Refuse Forever: Weaponizing Alignment for Injecting Bias in LLMs

Large Language Models (LLMs) are aligned to meet ethical standards and safety requirements by training them to refuse answering harmful or unsafe prompts. In this paper, we demonstrate how adversaries can exploit LLMs' alignment to implant bias, or enforce targeted censorship without degrading the model's responsiveness to unrelated topics. Specifically, we propose Subversive Alignment Injection (SAI), a poisoning attack that leverages the alignment mechanism to trigger refusal on specific topics or queries predefined by the adversary. Although it is perhaps not surprising that refusal can be induced through overalignment, we demonstrate how this refusal can be exploited to inject bias into the model. Surprisingly, SAI evades state-of-the-art poisoning defenses including LLM state forensics, as well as robust aggregation techniques that are designed to detect poisoning in FL settings. We demonstrate the practical dangers of this attack by illustrating its end-to-end impacts on LLM-powered application pipelines. For chat based applications such as ChatDoctor, with 1% data poisoning, the system refuses to answer healthcare questions to targeted racial category leading to high bias ($ΔDP$ of 23%). We also show that bias can be induced in other NLP tasks: for a resume selection pipeline aligned to refuse to summarize CVs from a selected university, high bias in selection ($ΔDP$ of 27%) results. Even higher bias ($ΔDP$~38%) results on 9 other chat based downstream applications.

cs.LG↗
arXiv2025

PRACtical: Subarray-Level Counter Update and Bank-Level Recovery Isolation for Efficient PRAC Rowhammer Mitigation

As DRAM density increases, Rowhammer becomes more severe due to heightened charge leakage, reducing the number of activations needed to induce bit flips. The DDR5 standard addresses this threat with in-DRAM per-row activation counters (PRAC) and the Alert Back-Off (ABO) signal to trigger mitigation. However, PRAC adds performance overhead by incrementing counters during the precharge phase, and recovery refreshes stalls the entire memory channel, even if only one bank is under attack. We propose PRACtical, a performance-optimized approach to PRAC+ABO that maintains the same security guarantees. First, we reduce counter update latency by introducing a centralized increment circuit, enabling overlap between counter updates and subsequent row activations in other subarrays. Second, we enhance the $RFM_{ab}$ mitigation by enabling bank-level granularity: instead of stalling the entire channel, only affected banks are paused. This is achieved through a DRAM-resident register that identifies attacked banks. PRACtical improves performance by 8% on average (up to 20%) over the state-of-the-art, reduces energy by 19%, and limits performance degradation from aggressive performance attacks to less than 6%, all while preserving Rowhammer protection.

cs.AR↗
arXiv2025

Layer-wise Alignment: Examining Safety Alignment Across Image Encoder Layers in Vision Language Models

Vision-language models (VLMs) have improved significantly in their capabilities, but their complex architecture makes their safety alignment challenging. In this paper, we reveal an uneven distribution of harmful information across the intermediate layers of the image encoder and show that skipping a certain set of layers and exiting early can increase the chance of the VLM generating harmful responses. We call it as "Image enCoder Early-exiT" based vulnerability (ICET). Our experiments across three VLMs: LLaVA-1.5, LLaVA-NeXT, and Llama 3.2, show that performing early exits from the image encoder significantly increases the likelihood of generating harmful outputs. To tackle this, we propose a simple yet effective modification of the Clipped-Proximal Policy Optimization (Clip-PPO) algorithm for performing layer-wise multi-modal RLHF for VLMs. We term this as Layer-Wise PPO (L-PPO). We evaluate our L-PPO algorithm across three multimodal datasets and show that it consistently reduces the harmfulness caused by early exits.

cs.CL↗
arXiv2025

I Know What You Sync: Covert and Side Channel Attacks on File Systems via syncfs

Operating Systems enforce logical isolation using abstractions such as processes, containers, and isolation technologies to protect a system from malicious or buggy code. In this paper, we show new types of side channels through the file system that break this logical isolation. The file system plays a critical role in the operating system, managing all I/O activities between the application layer and the physical storage device. We observe that the file system implementation is shared, leading to timing leakage when using common I/O system calls. Specifically, we found that modern operating systems take advantage of any flush operation (which saves cached blocks in memory to the SSD or disk) to flush all of the I/O buffers, even those used by other isolation domains. Thus, by measuring the delay of syncfs, the attacker can infer the I/O behavior of victim programs. We then demonstrate a syncfs covert channel attack on multiple file systems, including both Linux native file systems and the Windows file system, achieving a maximum bandwidth of 5 Kbps with an error rate of 0.15% on Linux and 7.6 Kbps with an error rate of 1.9% on Windows. In addition, we construct three side-channel attacks targeting both Linux and Android devices. On Linux devices, we implement a website fingerprinting attack and a video fingerprinting attack by tracking the write patterns of temporary buffering files. On Android devices, we design an application fingerprinting attack that leaks application write patterns during boot-up. The attacks achieve over 90% F1 score, precision, and recall. Finally, we demonstrate that these attacks can be exploited across containers implementing a container detection technique and a cross-container covert channel attack.

cs.CR↗
arXiv2025

Misaligned Roles, Misplaced Images: Structural Input Perturbations Expose Multimodal Alignment Blind Spots

Multimodal Language Models (MMLMs) typically undergo post-training alignment to prevent harmful content generation. However, these alignment stages focus primarily on the assistant role, leaving the user role unaligned, and stick to a fixed input prompt structure of special tokens, leaving the model vulnerable when inputs deviate from these expectations. We introduce Role-Modality Attacks (RMA), a novel class of adversarial attacks that exploit role confusion between the user and assistant and alter the position of the image token to elicit harmful outputs. Unlike existing attacks that modify query content, RMAs manipulate the input structure without altering the query itself. We systematically evaluate these attacks across multiple Vision Language Models (VLMs) on eight distinct settings, showing that they can be composed to create stronger adversarial prompts, as also evidenced by their increased projection in the negative refusal direction in the residual stream, a property observed in prior successful attacks. Finally, for mitigation, we propose an adversarial training approach that makes the model robust against input prompt perturbations. By training the model on a range of harmful and benign prompts all perturbed with different RMA settings, it loses its sensitivity to Role Confusion and Modality Manipulation attacks and is trained to only pay attention to the content of the query in the input prompt structure, effectively reducing Attack Success Rate (ASR) while preserving the model's general utility.

cs.CR↗
arXiv2025

NVBleed: Covert and Side-Channel Attacks on NVIDIA Multi-GPU Interconnect

Multi-GPU systems are becoming increasingly important in highperformance computing (HPC) and cloud infrastructure, providing acceleration for data-intensive applications, including machine learning workloads. These systems consist of multiple GPUs interconnected through high-speed networking links such as NVIDIA's NVLink. In this work, we explore whether the interconnect on such systems can offer a novel source of leakage, enabling new forms of covert and side-channel attacks. Specifically, we reverse engineer the operations of NVlink and identify two primary sources of leakage: timing variations due to contention and accessible performance counters that disclose communication patterns. The leakage is visible remotely and even across VM instances in the cloud, enabling potentially dangerous attacks. Building on these observations, we develop two types of covert-channel attacks across two GPUs, achieving a bandwidth of over 70 Kbps with an error rate of 4.78% for the contention channel. We develop two end-to-end crossGPU side-channel attacks: application fingerprinting (including 18 high-performance computing and deep learning applications) and 3D graphics character identification within Blender, a multi-GPU rendering application. These attacks are highly effective, achieving F1 scores of up to 97.78% and 91.56%, respectively. We also discover that leakage surprisingly occurs across Virtual Machines on the Google Cloud Platform (GCP) and demonstrate a side-channel attack on Blender, achieving F1 scores exceeding 88%. We also explore potential defenses such as managing access to counters and reducing the resolution of the clock to mitigate the two sources of leakage.

cs.CR↗
arXiv2025

Attention Eclipse: Manipulating Attention to Bypass LLM Safety-Alignment

Recent research has shown that carefully crafted jailbreak inputs can induce large language models to produce harmful outputs, despite safety measures such as alignment. It is important to anticipate the range of potential Jailbreak attacks to guide effective defenses and accurate assessment of model safety. In this paper, we present a new approach for generating highly effective Jailbreak attacks that manipulate the attention of the model to selectively strengthen or weaken attention among different parts of the prompt. By harnessing attention loss, we develop more effective jailbreak attacks, that are also transferrable. The attacks amplify the success rate of existing Jailbreak algorithms including GCG, AutoDAN, and ReNeLLM, while lowering their generation cost (for example, the amplified GCG attack achieves 91.2% ASR, vs. 67.9% for the original attack on Llama2-7B/AdvBench, using less than a third of the generation time).

cs.CR↗