Searcharxiv⌕ Search

arXiv subjects

Samuele Poppi

Publications and source records attributed to Samuele Poppi.

14 recordsLinked to original sources
arXiv2026

CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs

Large language models can reproduce memorized text verbatim, yet copyright defenses are usually evaluated under incompatible protocols. We introduce CopyShield, a controlled benchmark comparing three representative defenses at distinct intervention levels: contrastive decoding (output), Direct Preference Optimization (behavioral), and activation intervention (representation). We evaluate CopyShield on two model families, LLaMA-3.1-8B and Mistral-7B-v0.3, using controlled memorization over five public-domain books and a shared protocol measuring literal leakage, calibrated non-literal leakage, utility, and degeneracy. Across these methods, intervention level is associated with distinct compliance-utility trade-offs. On LLaMA-3.1-8B, contrastive decoding remains near-degeneracy-free (0-2%) but reaches a literal-suppression floor at NV-Recall 0.192-0.203. DPO nearly eliminates literal leakage (0.263 to 0.002) but induces paraphrase-loop degeneracy in 58% of QA outputs, with no utility gain over the SFT baseline. Activation intervention attains the lowest non-literal flagging rate (1/200) by blocking 84% of non-literal queries before generation. Human evaluation confirms that DPO has low coherence, whereas activation lowers perceived copyright risk through broad refusal. On Mistral-7B-v0.3, the output- and representation-level patterns persist, while DPO degeneracy falls to 10-14%, showing that its severity is model-dependent. Together, CopyShield provides cross-level reference baselines and identifies targeted non-literal suppression as an open challenge. The code is available at https://github.com/spotai-mbzuai/CopyShield.git.

cs.LG↗
arXiv2026

Context Inference Attacks Without Jailbreaks

Agentic AI systems are increasingly deployed to process sensitive data at inference time, such as healthcare records or financial documents assembled into a hidden \emph{context} before the system answers. Prior work has studied privacy risks primarily through \emph{jailbreaking} attacks that induce models to directly disclose sensitive content, but has largely overlooked the agentic setting where the context is assembled by the agent's own tool calls. We show that the agents we evaluate remain vulnerable to hidden-context leakage despite the controls we test against them, namely an instruction not to disclose the context, logit suppression, and context dilution. For instance, a web-browsing agent answering benign user queries still carries exploitable signals about records silently loaded into its context. We introduce and formalize \emph{context-inference attacks} through a security game and evaluate three settings under decreasing attacker knowledge and increasingly indirect delivery of the context: a known context, an unknown context, and a context the agent retrieves through its own tool calls. We distinguish a grey-box setting, in which the target model is used to score observations, from black-box settings in which the attacker scores with a surrogate it controls. We further characterize how leakage varies with query budget, context size, and target-model size. A single attack carries through all three settings without modification, reaching $100\%$ ASR on small candidate sets and $63\%$ at $1024$ candidates against a known context, $78.9$ AUROC when the template and surrounding records are unknown, $92.5$ AUROC when a 14B surrogate scores a 32B target, and $81.8$ AUROC when the records arrive as an agent's retrieval returns, against chance rates of $1/|\mathcal{Z}|$ and $50$ respectively.

cs.CR↗
arXiv2026

ReACT-CLIP: Response-Aware Test-Time Defense for Vision--Language Models

Training-free test-time defenses offer a practical way to improve the adversarial robustness of CLIP-style vision--language models without modifying the pretrained model. However, their correction strength is typically fixed for a narrow range of attack budgets, even though the attack budget is unknown at inference and the required correction varies across samples. We show that this mismatch causes existing defenses to degrade sharply as attacks strengthen. We introduce ReACT-CLIP, a response-conditioned test-time defense that separately determines how strongly each input should be corrected and whether defensive intervention is necessary. Our key observation is that the relative increase in CLIP visual-feature drift between low- and high-noise probes provides a graded, sample-specific proxy for correction demand. ReACT-CLIP maps this relative cross-noise drift to the Gaussian noise scale used to construct a stable, noise-averaged feature anchor, enabling the corrective reach to adapt to each input. To determine whether intervention is necessary, we further observe that clean inputs retain stable class-probability distributions under weak spatial augmentations, whereas adversarial inputs exhibit greater variation. ReACT-CLIP quantifies this variation using a prediction-instability score computed by Jensen--Shannon divergence and combines it with relative cross-noise drift to form the defensive intervention score. ReACT-CLIP requires no model or prompt training, and its correction-strength mapping is calibrated once and fixed across datasets and attack budgets. Across 12 downstream datasets, as well as ImageNet and its distribution-shifted variants, ReACT-CLIP delivers substantial robustness gains across diverse attack types and strengths while largely preserving clean accuracy.

cs.CV↗
arXiv2026

A Gravitational Interpretation of Fine-Tuning Reversion

Fine-tuning on harmless data can partially undo behaviors acquired earlier in training. Safety can erode under benign post-alignment updates, unlearned capabilities can re-emerge, latent traits can transfer through apparently unrelated supervision, and related post-alignment fragility appears in other generative settings. We argue these phenomena are usefully viewed through a common training-history lens. Our hypothesis is geometric: large early training phases create dominant behavioral manifolds, while later alignment or specialization phases are shallower displacements from them. Subsequent fine-tuning can therefore inherit a persistent reversion component pointing back toward a witness of the dominant manifold. We call this the gravitational interpretation of fine-tuning reversion. Across our main settings, representational drift rapidly acquires a component along a history-defined reversion direction (v_rev). In our main track, alignment with v_rev rises from cos = 0.429 +/- 0.052 after the first update to 0.647 +/- 0.021 by step 20. Across 24 run-step pairs, every observed alignment exceeds the p99 of an isotropic activation-space null. We demonstrate that selectively blocking motion along v_rev changes the final alignment at T=100 from 0.648 +/- 0.009 to -0.211 +/- 0.021 and reduces harmfulness from 19.0% +/- 4.0% to 8.5% +/- 1.5% with little task cost. These results support v_rev as a causally relevant mediator of early post-alignment reversion in our setup. Importantly, we do not claim that v_rev is the unique safety direction, nor that the dominant manifold is directly observed; rather, we identify a robust, history-defined direction that explains and partially controls early reversion dynamics.

cs.LG↗
arXiv2026

When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models

Failures in multi-turn reasoning models are largely invisible to terminal-score evaluation. A model can lock onto an unsafe stance early in a long dialogue, yet its final-turn refusal rate may appear indistinguishable from a robustly aligned baseline. To expose these hidden temporal dynamics, we propose a trace-level diagnostic - the CoT-Output 2x2 safety matrix. This framework labels every turn along two independent axes (internal reasoning and visible output), yielding four operationally defined failure cells: robust alignment, alignment faking, overt jailbreak, and a distinct failure mode we term context-injection failure (where the CoT maintains safe reasoning, but the visible output produces harm, highlighting a multi-turn manifestation of reasoning unfaithfulness). We evaluate three distilled reasoning targets against a fixed attacker across five oversight conditions, collecting 6750 turn-level observations on the Information-Hazard scenario. Our analysis reveals two reproducible vulnerabilities: an oversight paradox where explicit monitoring cues paradoxically increase alignment-faking rates rather than suppress them, and a context-injection failure where models lock onto unsafe external outputs despite safe internal states. We release the full dataset of multi-turn dialogues and CoT traces to support follow-up trace-diagnostic research.

cs.AI↗
arXiv2026

Adaptively Robust LLM Monitoring via Activation Watermarking

Providers monitor deployed large language models (LLMs) to detect misuse that they cannot prevent. LLM monitoring is deterministic and often openly available, so $\emph{adaptive}$ attackers with a local copy can search offline for prompts that elicit harmful behavior and evade detection. These attacks are especially concerning because providers never observe the misuse and cannot patch their defenses post-hoc. The core challenge is resisting adaptive attackers while preserving detection rates against non-adaptive ones. We propose $\emph{Activation Watermarking}$ (AWM), which randomizes monitoring through limited fine-tuning that aligns the LLM's hidden states with a secret key-derived direction whenever a response violates a policy. Detection is a similarity test on activations the provider already computes, and attackers who know everything but the key must optimize against differently keyed surrogate detectors. At a matched $1\%$ false-positive rate, such surrogate-based attacks evade every evaluated baseline at least $79\%$ of the time, but AWM more than halves this evasion rate. AWM also achieves the lowest evasion rate on three of four non-adaptive jailbreak families at a small drop in utility measured across seven benchmarks. Because random high-dimensional directions are near-orthogonal, assigning one watermark per policy enables attribution. Across 20 monitored policies, AWM identifies the violated policy with $80\%$ accuracy versus a $5\%$ chance baseline. AWM is (i) efficient, (ii) substantially more robust against adaptive attackers than related work and (iii) can attribute which policies were violated.

cs.CR↗
arXiv2025

Robust and Calibrated Detection of Authentic Multimedia Content

Generative models can synthesize highly realistic content, so-called deepfakes, that are already being misused at scale to undermine digital media authenticity. Current deepfake detection methods are unreliable for two reasons: (i) distinguishing inauthentic content post-hoc is often impossible (e.g., with memorized samples), leading to an unbounded false positive rate (FPR); and (ii) detection lacks robustness, as adversaries can adapt to known detectors with near-perfect accuracy using minimal computational resources. To address these limitations, we propose a resynthesis framework to determine if a sample is authentic or if its authenticity can be plausibly denied. We make two key contributions focusing on the high-precision, low-recall setting against efficient (i.e., compute-restricted) adversaries. First, we demonstrate that our calibrated resynthesis method is the most reliable approach for verifying authentic samples while maintaining controllable, low FPRs. Second, we show that our method achieves adversarial robustness against efficient adversaries, whereas prior methods are easily evaded under identical compute budgets. Our approach supports multiple modalities and leverages state-of-the-art inversion techniques.

cs.CV↗
arXiv2025

SPQR: A Multi-Dimensional Benchmark for Safety Alignment under Benign Model Adaptation

Text-to-image diffusion models can emit copyrighted, unsafe, or private content. Safety alignment aims to suppress specific concepts, yet evaluations seldom test whether safety persists under benign downstream fine-tuning routinely applied after deployment (e.g., LoRA personalization, style/domain adapters). We study the stability of current safety methods under benign fine-tuning and observe frequent breakdowns. As true safety alignment must withstand even benign post-deployment adaptations, we introduce the SPQR benchmark (Safety, Prompt adherence, Quality, and Robustness). SPQR is a single-scored metric that provides a unified, reproducible framework to evaluate how well safety-aligned diffusion models preserve safety, utility, and robustness under benign fine-tuning, by reporting a single leaderboard score to facilitate comparisons. We conduct multilingual, domain-specific, and out-of-distribution analyses, along with category-wise breakdowns, to identify when safety alignment fails after benign fine-tuning, ultimately showcasing SPQR as a concise yet comprehensive benchmark for T2I safety alignment techniques for T2I models.

cs.CR↗
arXiv2025

Mitigating Watermark Forgery in Generative Models via Randomized Key Selection

Watermarking enables GenAI providers to verify whether content was generated by their models. A watermark is a hidden signal in the content, whose presence can be detected using a secret watermark key. A core security threat are forgery attacks, where adversaries insert the provider's watermark into content \emph{not} produced by the provider, potentially damaging their reputation and undermining trust. Existing defenses resist forgery by embedding many watermarks with multiple keys into the same content, which can degrade model utility. However, forgery remains a threat when attackers can collect sufficiently many watermarked samples. We propose a defense that is provably forgery-resistant \emph{independent} of the number of watermarked content collected by the attacker, provided they cannot easily distinguish watermarks from different keys. Our scheme does not further degrade model utility. We randomize the watermark key selection for each query and accept content as genuine only if a watermark is detected by \emph{exactly} one key. We focus on the image and text modalities, but our defense is modality-agnostic, since it treats the underlying watermarking method as a black-box. Our method provably bounds the attacker's success rate and we empirically observe a reduction from near-perfect success rates to only $2\%$ at negligible computational overhead.

cs.CR↗
arXiv2025

Improving LLM First-Token Predictions in Multiple-Choice Question Answering via Output Prefilling

Large Language Models (LLMs) are increasingly evaluated on multiple-choice question answering (MCQA) tasks using *first-token probability* (FTP), which selects the answer option whose initial token has the highest likelihood. While efficient, FTP can be fragile: models may assign high probability to unrelated tokens (*misalignment*) or use a valid token merely as part of a generic preamble rather than as a clear answer choice (*misinterpretation*), undermining the reliability of symbolic evaluation. We propose a simple solution: the *prefilling attack*, a structured natural-language prefix (e.g., "*The correct option is:*") prepended to the model output. Originally explored in AI safety, we repurpose prefilling to steer the model to respond with a clean, valid option, without modifying its parameters. Empirically, the FTP with prefilling strategy substantially improves accuracy, calibration, and output consistency across a broad set of LLMs and MCQA benchmarks. It outperforms standard FTP and often matches the performance of open-ended generation approaches that require full decoding and external classifiers, while being significantly more efficient. Our findings suggest that prefilling is a simple, robust, and low-cost method to enhance the reliability of FTP-based evaluation in multiple-choice settings.

cs.CL↗
arXiv2024

Towards Understanding the Fragility of Multilingual LLMs against Fine-Tuning Attacks

Recent advancements in Large Language Models (LLMs) have sparked widespread concerns about their safety. Recent work demonstrates that safety alignment of LLMs can be easily removed by fine-tuning with a few adversarially chosen instruction-following examples, i.e., fine-tuning attacks. We take a further step to understand fine-tuning attacks in multilingual LLMs. We first discover cross-lingual generalization of fine-tuning attacks: using a few adversarially chosen instruction-following examples in one language, multilingual LLMs can also be easily compromised (e.g., multilingual LLMs fail to refuse harmful prompts in other languages). Motivated by this finding, we hypothesize that safety-related information is language-agnostic and propose a new method termed Safety Information Localization (SIL) to identify the safety-related information in the model parameter space. Through SIL, we validate this hypothesis and find that only changing 20% of weight parameters in fine-tuning attacks can break safety alignment across all languages. Furthermore, we provide evidence to the alternative pathways hypothesis for why freezing safety-related parameters does not prevent fine-tuning attacks, and we demonstrate that our attack vector can still jailbreak LLMs adapted to new languages.

cs.CL↗
arXiv2023

Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models

Large-scale vision-and-language models, such as CLIP, are typically trained on web-scale data, which can introduce inappropriate content and lead to the development of unsafe and biased behavior. This, in turn, hampers their applicability in sensitive and trustworthy contexts and could raise significant concerns in their adoption. Our research introduces a novel approach to enhancing the safety of vision-and-language models by diminishing their sensitivity to NSFW (not safe for work) inputs. In particular, our methodology seeks to sever "toxic" linguistic and visual concepts, unlearning the linkage between unsafe linguistic or visual items and unsafe regions of the embedding space. We show how this can be done by fine-tuning a CLIP model on synthetic data obtained from a large language model trained to convert between safe and unsafe sentences, and a text-to-image generator. We conduct extensive experiments on the resulting embedding space for cross-modal retrieval, text-to-image, and image-to-text generation, where we show that our model can be remarkably employed with pre-trained generative models. Our source code and trained models are available at: https://github.com/aimagelab/safe-clip.

cs.CV↗
arXiv2023

Multi-Class Unlearning for Image Classification via Weight Filtering

Machine Unlearning is an emerging paradigm for selectively removing the impact of training datapoints from a network. Unlike existing methods that target a limited subset or a single class, our framework unlearns all classes in a single round. We achieve this by modulating the network's components using memory matrices, enabling the network to demonstrate selective unlearning behavior for any class after training. By discovering weights that are specific to each class, our approach also recovers a representation of the classes which is explainable by design. We test the proposed framework on small- and medium-scale image classification datasets, with both convolution- and Transformer-based backbones, showcasing the potential for explainable solutions through unlearning.

cs.CV↗
arXiv2021

Revisiting The Evaluation of Class Activation Mapping for Explainability: A Novel Metric and Experimental Analysis

As the request for deep learning solutions increases, the need for explainability is even more fundamental. In this setting, particular attention has been given to visualization techniques, that try to attribute the right relevance to each input pixel with respect to the output of the network. In this paper, we focus on Class Activation Mapping (CAM) approaches, which provide an effective visualization by taking weighted averages of the activation maps. To enhance the evaluation and the reproducibility of such approaches, we propose a novel set of metrics to quantify explanation maps, which show better effectiveness and simplify comparisons between approaches. To evaluate the appropriateness of the proposal, we compare different CAM-based visualization methods on the entire ImageNet validation set, fostering proper comparisons and reproducibility.

cs.CV↗