arXiv subjects

Ihsen Alouani

Publications and source records attributed to Ihsen Alouani.

At least 19 recordsLinked to original sources

arXiv2026

Batch Normalization Amplifies Memorization and Privacy Risks

Batch Normalization (BN) is widely adopted to enable faster convergence and more stable training of deep neural networks. However, its impact on privacy and memorization has remained largely unexplored. In this work, we investigate the effect of BN layers on the memorization of atypical or outlier samples and its implications for privacy leakage. We conduct an extensive empirical study using three complementary approaches: (i) unintended memorization of out-of-distribution samples, (ii) per-sample influence, and (iii) susceptibility to membership inference attacks (MIA). Across multiple datasets and architectures, we consistently observe that BN substantially increases the memorization of outliers compared to models without BN. Critically, this amplified memorization translates directly into privacy vulnerabilities: models with BN exhibit significantly higher susceptibility to MIAs. We complement our empirical findings with a mechanistic analysis under the exact BN backward pass, which shows that BN amplifies the per-step margin growth of outlier samples during training. Our results highlight an underappreciated privacy risk associated with BN and provide both practical and theoretical insights into how normalization layers can amplify the influence of rare or sensitive training examples.

cs.LG↗

arXiv2026

Faults That Fortify: CNN Adversarial Robustness via GPU Undervolting

Convolutional Neural Networks (CNNs) face a dual challenge: vulnerability to adversarial attacks and prohibitive training cost. Adversarial training is effective but expensive, a burden that grows as learning shifts to the energy-constrained edge. This paper addresses both through GPU undervolting during training. Reducing supply voltage introduces stochastic perturbations that act as implicit regularization, improving robustness while lowering power. We characterize undervolting-induced faults at the bit level, then train LeNet, VGG-6, and MobileNetV3 on MNIST and CIFAR-10 under two training regimes, standard and adversarial, each at nominal and undervolted voltage, and evaluate all models against adversarial attacks. In both regimes, the undervolted model consistently achieves higher adversarial accuracy than its nominal-voltage counterpart, showing that hardware-induced faults strengthen even adversarial training. Because dynamic power scales quadratically with supply voltage, these robustness gains arrive with substantial energy savings. GPU undervolting is therefore a readily deployable hardware-level defense requiring no algorithmic change, and opens a promising direction in which robustness and energy efficiency move together.

cs.LG↗

arXiv2026

Emerging Threats and Countermeasures in Neuromorphic Systems: A Survey

Neuromorphic computing mimics brain-inspired mechanisms through spiking neurons and energy-efficient processing, offering a pathway to efficient in-memory computing (IMC). However, these advancements raise critical security and privacy concerns. As the adoption of bio-inspired architectures and memristive devices increases, so does the urgency to assess the vulnerability of these emerging technologies to hardware and software attacks. Emerging architectures introduce new attack surfaces, particularly due to asynchronous, event-driven processing and stochastic device behavior. The integration of memristors into neuromorphic hardware and software implementations in spiking neural networks offers diverse possibilities for advanced computing architectures, including their role in security-aware applications. This survey systematically analyzes the security landscape of neuromorphic systems, covering attack methodologies, side-channel vulnerabilities, and countermeasures. We focus on both hardware and software concerns relevant to spiking neural networks (SNNs) and hardware primitives, such as Physical Unclonable Functions (PUFs) and True Random Number Generators (TRNGs) for cryptographic and secure computation applications. We approach this analysis from diverse perspectives, from attack methodologies to countermeasure strategies that integrate efficiency and protection in brain-inspired hardware. This review not only maps the current landscape of security threats but provides a foundation for developing secure and trustworthy neuromorphic architectures.

cs.CR↗

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

Bypassing Prompt Injection Detectors through Evasive Injections

Large language models (LLMs) are increasingly used in interactive and retrieval-augmented systems, but they remain vulnerable to prompt injection attacks, where injected secondary prompts force the model to deviate from the user's instructions to execute a potentially malicious task defined by the adversary. Recent work shows that ML models trained on activation shifts from LLMs' hidden layers can detect such drift. In this paper, we demonstrate that these detectors are not robust to adaptive adversaries. We propose a multi-probe evasion attack that appends an adversarially optimised suffix to poisoned inputs, jointly optimising a universal suffix to simultaneously fool all layer-wise drift detectors while preserving the effectiveness of the underlying injection. Using a modified Greedy Coordinate Gradient (GCG) approach, we generate universal suffixes that make prompt injections consistently evasive across multiple probes simultaneously. On Phi-3 3.8B and Llama-3 8B, a single suffix achieves attack success rates of 93.91% and 99.63% in successfully evading all detectors simultaneously. These results show that activation-based task drift detectors are highly vulnerable to adaptive prompt injection attacks, motivating stronger defences against such threats. We also propose a defence based on adversarial suffix augmentation: we generate multiple suffixes, append one at random during forward passes, and train detectors on the resulting activations. This approach is found to be effective against evasive attacks.

cs.CR↗

arXiv2026

SalamahBench: Toward Standardized Safety Evaluation for Arabic Language Models

Safety alignment in Language Models (LMs) is fundamental for trustworthy AI. However, while different stakeholders are trying to leverage Arabic Language Models (ALMs), systematic safety evaluation of ALMs remains largely underexplored, limiting their mainstream uptake. Existing safety benchmarks and safeguard models are predominantly English-centric, limiting their applicability to Arabic Natural Language Processing (NLP) systems and obscuring fine-grained, category-level safety vulnerabilities. This paper introduces SalamaBench, a unified benchmark for evaluating the safety of ALMs, comprising $8,170$ prompts across $12$ different categories aligned with the MLCommons Safety Hazard Taxonomy. Constructed by harmonizing heterogeneous datasets through a rigorous pipeline involving AI filtering and multi-stage human verification, SalamaBench enables standardized, category-aware safety evaluation. Using this benchmark, we evaluate five state-of-the-art ALMs, including Fanar 1 and 2, ALLaM 2, Falcon H1R, and Jais 2, under multiple safeguard configurations, including individual guard models, majority-vote aggregation, and validation against human-annotated gold labels. Our results reveal substantial variation in safety alignment: while Fanar 2 achieves the lowest aggregate attack success rates, its robustness is uneven across specific harm domains. In contrast, Jais 2 consistently exhibits elevated vulnerability, indicating weaker intrinsic safety alignment. We further demonstrate that native ALMs perform substantially worse than dedicated safeguard models when acting as safety judges. Overall, our findings highlight the necessity of category-aware evaluation and specialized safeguard mechanisms for robust harm mitigation in ALMs.

cs.CL↗

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

Privacy in Federated Learning with Spiking Neural Networks

Spiking neural networks (SNNs) have emerged as prominent candidates for embedded and edge AI. Their inherent low power consumption makes them far more efficient than conventional ANNs in scenarios where energy budgets are tightly constrained. In parallel, federated learning (FL) has become the prevailing training paradigm in such settings, enabling on-device learning while limiting the exposure of raw data. However, gradient inversion attacks represent a critical privacy threat in FL, where sensitive training data can be reconstructed directly from shared gradients. While this vulnerability has been widely investigated in conventional ANNs, its implications for SNNs remain largely unexplored. In this work, we present the first comprehensive empirical study of gradient leakage in SNNs across diverse data domains. SNNs are inherently non-differentiable and are typically trained using surrogate gradients, which we hypothesized would be less correlated with the original input and thus less informative from a privacy perspective. To investigate this, we adapt different gradient leakage attacks to the spike domain. Our experiments reveal a striking contrast with conventional ANNs: whereas ANN gradients reliably expose salient input content, SNN gradients yield noisy, temporally inconsistent reconstructions that fail to recover meaningful spatial or temporal structure. These results indicate that the combination of event-driven dynamics and surrogate-gradient training substantially reduces gradient informativeness. To the best of our knowledge, this work provides the first systematic benchmark of gradient inversion attacks for spiking architectures, highlighting the inherent privacy-preserving potential of neuromorphic computation.

cs.LG↗

arXiv2025

Stealth by Conformity: Evading Robust Aggregation through Adaptive Poisoning

Federated Learning (FL) is a distributed learning paradigm designed to address privacy concerns. However, FL is vulnerable to poisoning attacks, where Byzantine clients compromise the integrity of the global model by submitting malicious updates. Robust aggregation methods have been widely adopted to mitigate such threats, relying on the core assumption that malicious updates are inherently out-of-distribution and can therefore be identified and excluded before aggregating client updates. In this paper, we challenge this underlying assumption by showing that a model can be poisoned while keeping malicious updates within the main distribution. We propose Chameleon Poisoning (CHAMP), an adaptive and evasive poisoning strategy that exploits side-channel feedback from the aggregation process to guide the attack. Specifically, the adversary continuously infers whether its malicious contribution has been incorporated into the global model and adapts accordingly. This enables a dynamic adjustment of the local loss function, balancing a malicious component with a camouflaging component, thereby increasing the effectiveness of the poisoning while evading robust aggregation defenses. CHAMP enables more effective and evasive poisoning, highlighting a fundamental limitation of existing robust aggregation defenses and underscoring the need for new strategies to secure federated learning against sophisticated adversaries. Our approach is evaluated in two datasets reaching an average increase of 47.07% in attack success rate against nine robust aggregation defenses.

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

On Jailbreaking Quantized Language Models Through Fault Injection Attacks

The safety alignment of Language Models (LMs) is a critical concern, yet their integrity can be challenged by direct parameter manipulation attacks, such as those potentially induced by fault injection. As LMs are increasingly deployed using low-precision quantization for efficiency, this paper investigates the efficacy of such attacks for jailbreaking aligned LMs across different quantization schemes. We propose gradient-guided attacks, including a tailored progressive bit-level search algorithm introduced herein and a comparative word-level (single weight update) attack. Our evaluation on Llama-3.2-3B, Phi-4-mini, and Llama-3-8B across FP16 (baseline), and weight-only quantization (FP8, INT8, INT4) reveals that quantization significantly influences attack success. While attacks readily achieve high success (>80% Attack Success Rate, ASR) on FP16 models, within an attack budget of 25 perturbations, FP8 and INT8 models exhibit ASRs below 20% and 50%, respectively. Increasing the perturbation budget up to 150 bit-flips, FP8 models maintained ASR below 65%, demonstrating some resilience compared to INT8 and INT4 models that have high ASR. In addition, analysis of perturbation locations revealed differing architectural targets across quantization schemes, with (FP16, INT4) and (INT8, FP8) showing similar characteristics. Besides, jailbreaks induced in FP16 models were highly transferable to subsequent FP8/INT8 quantization (<5% ASR difference), though INT4 significantly reduced transferred ASR (avg. 35% drop). These findings highlight that while common quantization schemes, particularly FP8, increase the difficulty of direct parameter manipulation jailbreaks, vulnerabilities can still persist, especially through post-attack quantization.

cs.CR↗

arXiv2025

SnatchML: Hijacking ML models without Training Access

Model hijacking can cause significant accountability and security risks since the owner of a hijacked model can be framed for having their model offer illegal or unethical services. Prior works consider model hijacking as a training time attack, whereby an adversary requires full access to the ML model training. In this paper, we consider a stronger threat model for an inference-time hijacking attack, where the adversary has no access to the training phase of the victim model. Our intuition is that ML models, which are typically over-parameterized, might have the capacity to (unintentionally) learn more than the intended task they are trained for. We propose SnatchML, a new training-free model hijacking attack, that leverages the extra capacity learnt by the victim model to infer different tasks that can be semantically related or unrelated to the original one. Our results on models deployed on AWS Sagemaker showed that SnatchML can deliver high accuracy on hijacking tasks. Interestingly, while all previous approaches are limited by the number of classes in the benign task, SnatchML can hijack models for tasks that contain more classes than the original. We explore different methods to mitigate this risk; We propose meta-unlearning, which is designed to help the model unlearn a potentially malicious task while training for the original task. We also provide insights on over-parametrization as a possible inherent factor that facilitates model hijacking, and accordingly, we propose a compression-based countermeasure to counteract this attack. We believe this work offers a previously overlooked perspective on model hijacking attacks, presenting a stronger threat model and higher applicability in real-world contexts.

cs.CR↗

arXiv2025

Input-Triggered Hardware Trojan Attack on Spiking Neural Networks

Neuromorphic computing based on spiking neural networks (SNNs) is emerging as a promising alternative to traditional artificial neural networks (ANNs), offering unique advantages in terms of low power consumption. However, the security aspect of SNNs is under-explored compared to their ANN counterparts. As the increasing reliance on AI systems comes with unique security risks and challenges, understanding the vulnerabilities and threat landscape is essential as neuromorphic computing matures. In this effort, we propose a novel input-triggered Hardware Trojan (HT) attack for SNNs. The HT mechanism is condensed in the area of one neuron. The trigger mechanism is an input message crafted in the spiking domain such that a selected neuron produces a malicious spike train that is not met in normal settings. This spike train triggers a malicious modification in the neuron that forces it to saturate, firing permanently and failing to recover to its resting state even when the input activity stops. The excessive spikes pollute the network and produce misleading decisions. We propose a methodology to select an appropriate neuron and to generate the input pattern that triggers the HT payload. The attack is illustrated by simulation on three popular benchmarks in the neuromorphic community. We also propose a hardware implementation for an analog spiking neuron and a digital SNN accelerator, demonstrating that the HT has a negligible area and power footprint and, thereby, can easily evade detection.

cs.NE↗

arXiv2025

Mind the Gap: Detecting Black-box Adversarial Attacks in the Making through Query Update Analysis

Adversarial attacks remain a significant threat that can jeopardize the integrity of Machine Learning (ML) models. In particular, query-based black-box attacks can generate malicious noise without having access to the victim model's architecture, making them practical in real-world contexts. The community has proposed several defenses against adversarial attacks, only to be broken by more advanced and adaptive attack strategies. In this paper, we propose a framework that detects if an adversarial noise instance is being generated. Unlike existing stateful defenses that detect adversarial noise generation by monitoring the input space, our approach learns adversarial patterns in the input update similarity space. In fact, we propose to observe a new metric called Delta Similarity (DS), which we show it captures more efficiently the adversarial behavior. We evaluate our approach against 8 state-of-the-art attacks, including adaptive attacks, where the adversary is aware of the defense and tries to evade detection. We find that our approach is significantly more robust than existing defenses both in terms of specificity and sensitivity.

cs.CR↗

arXiv2025

Are Neuromorphic Architectures Inherently Privacy-preserving? An Exploratory Study

While machine learning (ML) models are becoming mainstream, especially in sensitive application areas, the risk of data leakage has become a growing concern. Attacks like membership inference (MIA) have shown that trained models can reveal sensitive data, jeopardizing confidentiality. While traditional Artificial Neural Networks (ANNs) dominate ML applications, neuromorphic architectures, specifically Spiking Neural Networks (SNNs), are emerging as promising alternatives due to their low power consumption and event-driven processing, akin to biological neurons. Privacy in ANNs is well-studied; however, little work has explored the privacy-preserving properties of SNNs. This paper examines whether SNNs inherently offer better privacy. Using MIAs, we assess the privacy resilience of SNNs versus ANNs across diverse datasets. We analyze the impact of learning algorithms (surrogate gradient and evolutionary), frameworks (snnTorch, TENNLab, LAVA), and parameters on SNN privacy. Our findings show that SNNs consistently outperform ANNs in privacy preservation, with evolutionary algorithms offering additional resilience. For instance, on CIFAR-10, SNNs achieve an AUC of 0.59, significantly lower than ANNs' 0.82, and on CIFAR-100, SNNs maintain an AUC of 0.58 compared to ANNs' 0.88. Additionally, we explore the privacy-utility trade-off with Differentially Private Stochastic Gradient Descent (DPSGD), finding that SNNs sustain less accuracy loss than ANNs under similar privacy constraints.

cs.LG↗

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↗

arXiv2024

APARATE: Adaptive Adversarial Patch for CNN-based Monocular Depth Estimation for Autonomous Navigation

In recent times, monocular depth estimation (MDE) has experienced significant advancements in performance, largely attributed to the integration of innovative architectures, i.e., convolutional neural networks (CNNs) and Transformers. Nevertheless, the susceptibility of these models to adversarial attacks has emerged as a noteworthy concern, especially in domains where safety and security are paramount. This concern holds particular weight for MDE due to its critical role in applications like autonomous driving and robotic navigation, where accurate scene understanding is pivotal. To assess the vulnerability of CNN-based depth prediction methods, recent work tries to design adversarial patches against MDE. However, the existing approaches fall short of inducing a comprehensive and substantially disruptive impact on the vision system. Instead, their influence is partial and confined to specific local areas. These methods lead to erroneous depth predictions only within the overlapping region with the input image, without considering the characteristics of the target object, such as its size, shape, and position. In this paper, we introduce a novel adversarial patch named APARATE. This patch possesses the ability to selectively undermine MDE in two distinct ways: by distorting the estimated distances or by creating the illusion of an object disappearing from the perspective of the autonomous system. Notably, APARATE is designed to be sensitive to the shape and scale of the target object, and its influence extends beyond immediate proximity. APARATE, results in a mean depth estimation error surpassing $0.5$, significantly impacting as much as $99\%$ of the targeted region when applied to CNN-based MDE models. Furthermore, it yields a significant error of $0.34$ and exerts substantial influence over $94\%$ of the target region in the context of Transformer-based MDE.

cs.CV↗

arXiv2024

SSAP: A Shape-Sensitive Adversarial Patch for Comprehensive Disruption of Monocular Depth Estimation in Autonomous Navigation Applications

Monocular depth estimation (MDE) has advanced significantly, primarily through the integration of convolutional neural networks (CNNs) and more recently, Transformers. However, concerns about their susceptibility to adversarial attacks have emerged, especially in safety-critical domains like autonomous driving and robotic navigation. Existing approaches for assessing CNN-based depth prediction methods have fallen short in inducing comprehensive disruptions to the vision system, often limited to specific local areas. In this paper, we introduce SSAP (Shape-Sensitive Adversarial Patch), a novel approach designed to comprehensively disrupt monocular depth estimation (MDE) in autonomous navigation applications. Our patch is crafted to selectively undermine MDE in two distinct ways: by distorting estimated distances or by creating the illusion of an object disappearing from the system's perspective. Notably, our patch is shape-sensitive, meaning it considers the specific shape and scale of the target object, thereby extending its influence beyond immediate proximity. Furthermore, our patch is trained to effectively address different scales and distances from the camera. Experimental results demonstrate that our approach induces a mean depth estimation error surpassing 0.5, impacting up to 99% of the targeted region for CNN-based MDE models. Additionally, we investigate the vulnerability of Transformer-based MDE models to patch-based attacks, revealing that SSAP yields a significant error of 0.59 and exerts substantial influence over 99% of the target region on these models.

cs.CV↗