论文检索

Social Aspects · Trustworthy Machine Learning

Zhewen Yao, Yao Zhu, Shiliang Zhang

Existing adversarial attacks on Large Vision-Language Models (LVLMs) often struggle with limited transferability to black-box models or produce perceptible artifacts that are easily detected. This paper presents Progressive Semantic Infusion (PSI), a diffusion-based attack that progressively aligns and infuses natural target semantics. To improve transferability, PSI leverages diffusion priors to better align adversarial examples with the natural image distribution and employs progressive alignment to mitigate overfitting on a single fixed surrogate objective. To enhance stealthiness, PSI embeds source-aware cues during denoising to preserve visual fidelity and avoid detectable artifacts. Experiments show that PSI effectively attacks open-source, adversarially trained, and commercial VLMs, including GPT-5 and Grok-4, surpassing existing methods in both transferability and stealthiness. Our findings highlight a critical vulnerability in modern vision-language systems and offer valuable insights towards building more robust and trustworthy multimodal models.

Social Aspects · Fairness, Equity, Justice and Safety

Vincent Siu, Nathan Henry, Nicholas Crispino, Yang Liu, Dawn Song, Chenguang Wang

Current safety evaluations of language models rely on benchmark-based assessments that may miss targeted vulnerabilities. We present RepIt, a simple and data-efficient framework for isolating concept-specific representations in LM activations. While existing steering methods already achieve high attack success rates through broad interventions, RepIt enables a more concerning capability: selective suppression of refusal on targeted concepts while preserving refusal elsewhere. Across five frontier LMs, RepIt produces evaluation-evading models that answer questions related to weapons of mass destruction while still scoring as safe on standard benchmarks. We find the edit of the steering vector localizes to just 100-200 coordinates, and robust concept vectors can be extracted from as few as a dozen examples on a single A6000, highlighting how targeted, hard-to-detect modifications can exploit evaluation blind spots with minimal resources. By demonstrating precise concept disentanglement, this work exposes critical vulnerabilities in current safety evaluation practices and demonstrates an immediate need for more comprehensive, representation-aware assessments.

Social Aspects · Accountability, Transparency and Interpretability

Oscar Hill, Mateo Espinosa Zarlenga, Mateja Jamnik

Modern deep neural networks remain challenging to interpret due to the opacity of their latent representations, impeding model understanding, debugging, and debiasing. Concept Embedding Models (CEMs) address this by mapping inputs to human-interpretable concept representations from which tasks can be predicted. Yet, CEMs fail to represent inter-concept relationships and require concept annotations at different granularities during training, limiting their applicability. In this paper, we introduce *Hierarchical Concept Embedding Models* (HiCEMs), a new family of CEMs that explicitly model concept relationships through hierarchical structures. To enable HiCEMs in real-world settings, we propose *Concept Splitting*, a method for automatically discovering finer-grained sub-concepts from a pretrained CEM’s embedding space without requiring additional annotations. This allows HiCEMs to generate fine-grained explanations from limited concept labels, reducing annotation burdens. Our evaluation across multiple datasets, including a user study and experiments on *PseudoKitchens*, a newly proposed concept-based dataset of 3D kitchen renders, demonstrates that (1) Concept Splitting discovers human-interpretable sub-concepts absent during training that can be used to train highly accurate HiCEMs, and (2) HiCEMs enable powerful test-time concept interventions at different granularities, leading to improved task accuracy.

Social Aspects · Trustworthy Machine Learning

Justin Lee, Zheda Mai, Jinsu Yoo, Chongyu Fan, Cheng Zhang, Wei-Lun Chao

Machine unlearning—the ability to remove designated concepts from a pre-trained model—has advanced rapidly, particularly for text-to-image diffusion models. However, existing methods typically assume that unlearning requests arrive all at once, whereas in practice they often arrive sequentially. We present the first systematic study of continual unlearning in text-to-image diffusion models and show that popular unlearning methods suffer from rapid utility collapse: after only a few requests, models forget retained knowledge and generate degraded images. We trace this failure to cumulative parameter drift from the pre-training weights and argue that regularization is crucial to addressing it. To this end, we study a suite of add-on regularizers that (1) mitigate drift and (2) remain compatible with existing unlearning methods. Beyond generic regularizers, we show that semantic awareness is essential for preserving concepts close to the unlearning target, and propose a gradient-projection method that constrains parameter drift orthogonal to their subspace. This substantially improves continual unlearning performance and is complementary to other regularizers for further gains. Taken together, our study establishes continual unlearning as a fundamental challenge in text-to-image generation and provides insights, baselines, and open directions for advancing safe and accountable generative AI.

Social Aspects · Fairness, Equity, Justice and Safety

Alexander Panfilov, Evgenii Kortukov, Kristina Nikolić, Matthias Bethge, Sebastian Lapuschkin, Wojciech Samek, Ameya Prabhu, Maksym Andriushchenko, Jonas Geiping

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for \textit{dishonesty} as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are crafted to be subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool \emph{all} output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a \emph{honeypot} against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using them as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

Social Aspects · Fairness, Equity, Justice and Safety

Mikhail Terekhov, Alexander Panfilov, Daniil Dzenhaliou, Caglar Gulcehre, Maksym Andriushchenko, Ameya Prabhu, Jonas Geiping

AI control protocols serve as a defense mechanism to stop untrusted LLM agents from causing harm in autonomous settings. Prior work treats this as a security problem, stress testing with exploits that use the deployment context to subtly complete harmful side tasks, such as backdoor insertion. In practice, most AI control protocols are fundamentally based on LLM monitors, which can become a central point of failure. We study \textit{adaptive} attacks by an untrusted model that knows the protocol and the monitor model, which is plausible if the untrusted model was trained with a later knowledge cutoff or can search for this information autonomously. We instantiate a simple adaptive attack vector by which the attacker embeds known or zero-shot prompt injections in the model outputs. Using this tactic, frontier models consistently evade diverse monitors and complete malicious tasks on two main AI control benchmarks. The attack works universally against current protocols that rely on a monitor. Furthermore, the recent Defer-to-Resample protocol even backfires, as its resampling amplifies the prompt injection and effectively reframes it as a best-of-$n$ attack. In general, adaptive attacks on monitor models represent a major blind spot in current control protocols and should become a standard component of evaluations for future AI control mechanisms.

Social Aspects · Privacy-preserving Statistics and Machine Learning

Aviraj Newatia, Michael Cooper, Viet Nguyen, Rahul G. Krishnan

Training machine learning models requires the storage of large datasets, which often contain sensitive or private data. Storing data is associated with a number of potential risks which increase over time, such as database breaches and malicious adversaries. Machine unlearning is the study of methods to efficiently remove the influence of training data subsets from previously-trained models. Existing unlearning methods typically require direct access to the "forget set"---the data to be forgotten-and organisations must retain this data for unlearning rather than deleting it immediately upon request, increasing risks associated with the forget set. We introduce partially-blind unlearning---utilizing auxiliary information to unlearn without explicit access to the forget set. We also propose a practical framework Reload, a partially-blind method based on gradient optimization and structured weight sparsification to operationalize partially-blind unlearning. We show that Reload efficiently unlearns, approximating models retrained from scratch, and outperforms several forget set-dependent approaches. On language models, Reload unlearns entities using <0.025\% of the retain set and <7\% of model weights in <8 minutes on Llama2-7B. In the corrective case, Reload achieves unlearning even when only 10\% of corrupted data is identified.

Social Aspects · Accountability, Transparency and Interpretability

Nikhil Prakash, Natalie Shapira, Arnab Sen Sharma, Christoph Riedl, Yonatan Belinkov, Tamar Shaham, David Bau, Atticus Geiger

How do language models (LMs) represent characters’ beliefs, especially when those beliefs may differ from reality? This question lies at the heart of understanding the Theory of Mind (ToM) capabilities of LMs. We analyze LMs' ability to reason about characters’ beliefs using causal mediation and abstraction. We construct a dataset, CausalToM, consisting of simple stories where two characters independently change the state of two objects, potentially unaware of each other's actions. Our investigation uncovered a pervasive algorithmic pattern that we call a lookback mechanism, which enables the LM to recall important information when it becomes necessary. The LM binds each character-object-state triple together by co-locating their reference information, represented as Ordering IDs (OIs), in low-rank subspaces of the state token's residual stream. When asked about a character's beliefs regarding the state of an object, the binding lookback retrieves the correct state OI and then the answer lookback retrieves the corresponding state token. When we introduce text specifying that one character is (not) visible to the other, we find that the LM first generates a visibility ID encoding the relation between the observing and the observed character OIs. In a visibility lookback, this ID is used to retrieve information about the observed character and update the observing character's beliefs. Our work provides insights into belief tracking mechanisms, taking a step toward reverse-engineering ToM reasoning in LMs.

Social Aspects · Accountability, Transparency and Interpretability

Arnab Sen Sharma, Giordano Rogers, Natalie Shapira, David Bau

We investigate the mechanisms underlying a range of list-processing tasks in LLMs, and we find that they have learned to encode a compact, causal representation of a general filtering operation that mirrors the generic ``filter'' function of functional programming. Using causal mediation analysis on a diverse set of list-processing tasks, we find that a small number of attention heads, which we dub *filter heads*, encode a compact representation of the filtering predicate in their query states at certain tokens. We demonstrate that this predicate representation is general and portable: it can be extracted and reapplied to execute the same filtering operation on different collections, presented in different formats, languages, or even in tasks. However, we also identify situations where LMs can exploit a different strategy for filtering: eagerly evaluating if an item satisfies the predicate and storing this intermediate result as a flag directly in the item representations. Our results reveal that transformer LMs can develop human-interpretable implementations of abstract computational operations that generalize in ways that are surprisingly similar to strategies used in traditional functional programming patterns.

Social Aspects · Fairness, Equity, Justice and Safety

Hyunjun Jeon, Kyuyoung Kim, Jinwoo Shin

Modern language models can readily extract sensitive information from unstructured text, making redaction—the selective removal of such information—critical for data security. However, existing benchmarks for redaction typically focus on predefined categories of data such as personally identifiable information (PII) or evaluate specific techniques like masking. To address this limitation, we introduce RedacBench, a comprehensive benchmark for evaluating policy-conditioned redaction across domains and strategies. Constructed from 514 human-authored texts spanning individual, corporate, and government sources, paired with 187 security policies, RedacBench measures a model's ability to selectively remove policy-violating information while preserving the original semantics. We quantify performance using 8,053 annotated propositions that capture all inferable information in each text. This enables assessment of both security—the removal of sensitive propositions—and utility—the preservation of non-sensitive propositions. Experiments across multiple redaction strategies and state-of-the-art language models show that while more advanced models can improve security, preserving utility remains a challenge. To facilitate future research, we release RedacBench along with a web-based playground for dataset customization and evaluation. Available at https://hyunjunian.github.io/redaction-playground/.

Social Aspects · Fairness, Equity, Justice and Safety

Wen Huang, Jiarui Yang, Tao Dai, Jiawei Li, Shaoxiong Zhan, Bin Wang, Shu-Tao Xia

Visual manipulation localization (VML) aims to identify tampered regions in images and videos, a task that has become increasingly challenging with the rise of advanced editing tools. Existing methods face two central issues. The first is resolution diversity. Resizing or padding can distort subtle forensic cues and introduce unnecessary computational cost. The second is the difficulty of extending spatial models for images to spatio-temporal inputs in videos, which often results in maintaining separate architectures for the two data types. To address these challenges, we propose RelayFormer, a unified framework that adapts to varying resolutions and naturally handles both static and temporal visual data. RelayFormer partitions inputs into fixed-size sub-images and introduces Global Local Relay (GLR) tokens that propagate structured context through a relay-based attention mechanism. This design enables efficient exchange of global cues, such as semantic or temporal consistency, while preserving fine-grained manipulation artifacts. Unlike prior approaches that depend on uniform resizing or sparse attention, RelayFormer scales to variable resolutions and video sequences with minimal overhead. Experiments across diverse benchmarks demonstrate superior performance and strong efficiency, combining resolution adaptivity without interpolation or excessive padding, unified processing for images and videos, and a favorable balance between accuracy and computational cost. Code is available at~\href{https://github.com/WenOOI/RelayFormer}{https://github.com/WenOOI/RelayFormer}.

Social Aspects · Privacy-preserving Statistics and Machine Learning

Aurélien Bellet, Edwige Cyffers, Davide Frey, Romaric Gaudel, Dimitri Lerévérend, Francois Taiani

Decentralized Learning (DL) enables users to collaboratively train models without sharing raw data by iteratively averaging local updates with neighbors in a network graph. This setting is increasingly popular for its scalability and its ability to keep data local under user control. Strong privacy guarantees in DL are typically achieved through Differential Privacy (DP), with results showing that DL can even amplify privacy by disseminating noise across peer-to-peer communications. Yet in practice, the observed privacy-utility trade-off often appears worse than in centralized training, which may be due to limitations in current DP accounting methods for DL. In this paper, we show that recent advances in centralized DP accounting based on Matrix Factorization (MF) for analyzing temporal noise correlations can also be leveraged in DL. By generalizing existing MF results, we show how to cast both standard DL algorithms and common trust models into a unified formulation. This yields tighter privacy accounting for existing DP-DL algorithms and provides a principled way to develop new ones. To demonstrate the approach, we introduce MAFALDA-SGD, a gossip-based DL algorithm with user-level correlated noise that outperforms existing methods on synthetic and real-world graphs.

Social Aspects · Trustworthy Machine Learning

Thibaud Gloaguen, Mark Vero, Robin Staab, Martin Vechev

Finetuning open-weight Large Language Models (LLMs) is standard practice for achieving task-specific performance improvements. Until now, finetuning has been regarded as a controlled and secure process in which training on benign datasets leads to predictable behaviors. In this paper, we demonstrate, for the first time, that an adversary can create compromised LLMs that are performant and benign, yet exhibit adversarial behaviors once finetuned by downstream users. To this end, we propose an attack, FAB (Finetuning-activated Adversarial Behaviors), which compromises an LLM via meta-learning techniques that simulate downstream finetuning, explicitly optimizing for the emergence of adversarial behaviors in the finetuned models. At the same time, the compromised LLM is regularized to retain general capabilities and to exhibit no adversarial behaviors prior to finetuning. As a result, when users finetune (e.g., instruction-tuning, distillation, DPO) the seemingly benign model on their own datasets, they unknowingly trigger its dormant adversarial behavior. We experimentally demonstrate the effectiveness of FAB across multiple LLMs and three commonly considered target behaviors: unsolicited advertising, jailbreakability, and over-refusal. We show that FAB-triggers are robust to various finetuning choices made by the user (e.g., dataset, number of steps, scheduler, post-training algorithm). Our findings challenge prevailing assumptions on the security of finetuning, revealing a critical attack vector.

Social Aspects · Trustworthy Machine Learning

Mengxuan Hu, Vivek Datla, Anoop Kumar, Zihan Guan, Sheng Li, Alfy Samuel, Daben Liu

Recent advances in alignment techniques such as Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Direct Preference Optimization (DPO) have improved the safety of large language models (LLMs). However, these LLMs remain vulnerable to jailbreak attacks that disguise harmful intent through indirect or deceptive phrasing. Using causal intervention, we empirically demonstrate that this vulnerability stems from shallow alignment mechanisms that lack deep reasoning, often rejecting harmful prompts without truly understanding why they are harmful. To mitigate this vulnerability, we propose enhancing alignment through reasoning-aware post-training. We construct and release a novel Chain-of-Thought (CoT) fine-tuning dataset that includes both utility-oriented and safety-critical prompts with step-by-step rationales. Fine-tuning on this dataset encourages models to produce principled refusals grounded in reasoning, outperforming standard SFT baselines. Furthermore, inspired by failure patterns in CoT fine-tuning, we introduce **Alignment-Weighted DPO**, which targets the most problematic parts of an output by assigning different preference weights to the reasoning and final-answer segments. This produces finer-grained, targeted updates than vanilla DPO and improves robustness to diverse jailbreak strategies. Extensive experiments across multiple safety and utility benchmarks show that our method consistently improves alignment robustness while maintaining overall model utility.

Social Aspects · Accountability, Transparency and Interpretability

Thibaud Gloaguen, Robin Staab, Nikola Jovanović, Martin Vechev

We introduce the first watermark tailored for diffusion language models (DLMs), an emergent LLM paradigm able to generate tokens in arbitrary order, in contrast to standard autoregressive language models (ARLMs) which generate tokens sequentially. While there has been much work in ARLM watermarking, a key challenge when attempting to apply these schemes directly to the DLM setting is that they rely on previously generated tokens, which are not always available with DLM generation. In this work we address this challenge by: (i) applying the watermark in expectation over the context even when some context tokens are yet to be determined, and (ii) promoting tokens which increase the watermark strength when used as context for other tokens. This is accomplished while keeping the watermark detector unchanged. Our experimental evaluation demonstrates that the DLM watermark leads to a >99\% true positive rate with minimal quality impact and achieves similar robustness to existing ARLM watermarks, enabling for the first time reliable DLM watermarking.

Social Aspects · Accountability, Transparency and Interpretability

Thibaud Gloaguen, Robin Staab, Nikola Jovanović, Martin Vechev

Most LLM fingerprinting methods teach the model to respond to a few fixed queries with predefined atypical responses (keys). This memorization often does not survive common deployment steps such as finetuning or quantization, and such keys can be easily detected and filtered from LLM responses, ultimately breaking the fingerprint. To overcome these limitations we introduce *LLM fingerprinting via semantically conditioned watermarks*, replacing fixed query sets with a broad semantic domain, and replacing brittle atypical keys with a statistical watermarking signal diffused throughout each response. After teaching the model to watermark its responses only to prompts from a predetermined domain e.g., French language, the model owner can use queries from that domain to reliably detect the fingerprint and verify ownership. As we confirm in our thorough experimental evaluation, our fingerprint is both stealthy and robust to all common deployment scenarios.

Social Aspects · Trustworthy Machine Learning

Kazuki Egashira, Robin Staab, Thibaud Gloaguen, Mark Vero, Martin Vechev

Model pruning, i.e., removing a subset of model weights, has become a prominent approach to reducing the memory footprint of large language models (LLMs) during inference. Notably, popular inference engines, such as vLLM, enable users to conveniently prune downloaded models before they are deployed. While the utility and efficiency of pruning methods have improved significantly, the security implications of pruning remain underexplored. In this work, for the first time, we show that modern LLM pruning methods can be maliciously exploited. In particular, an adversary can construct a model that appears benign yet, once pruned, exhibits malicious behaviors. Our method is based on the idea that the adversary can compute a proxy metric that estimates how likely each parameter is to be pruned. With this information, the adversary can first inject a malicious behavior into those parameters that are unlikely to be pruned. Then, they can repair the model by using parameters that are \textit{likely} to be pruned, effectively canceling out the injected behavior in the unpruned model. We demonstrate the severity of our attack through extensive evaluation on five models; after any of the pruning in vLLM are applied (Magnitude, Wanda, and SparseGPT), it consistently exhibits strong malicious behaviors in a diverse set of attack scenarios (success rates of up to 95.7% for jailbreak, 98.7% for benign instruction refusal, and 99.5% for targeted content injection). Our results reveal a critical deployment-time security gap and underscore the urgent need for stronger security awareness in model compression.

Social Aspects · Trustworthy Machine Learning

Pedro Ferreira, Wilker Aziz, Ivan Titov

Chain-of-thought explanations are widely used to inspect the decision process of large language models (LLMs) and to evaluate the trustworthiness of model outputs, making them important for effective collaboration between LLMs and humans. We demonstrate that preference optimization -- a key step in the alignment phase -- can inadvertently reduce the faithfulness of these explanations. This occurs because the reward model (RM), which guides alignment, is tasked with optimizing both the expected quality of the response and the appropriateness of the explanations (e.g., minimizing bias or adhering to safety standards), creating potential conflicts. The RM lacks a mechanism to assess the consistency between the model’s internal decision process and the generated explanation. Consequently, the LLM may engage in ``reward hacking'' by producing a final response that scores highly while giving an explanation tailored to maximize reward rather than accurately reflecting its reasoning. To address this issue, we propose enriching the RM’s input with a causal attribution of the prediction, allowing the RM to detect discrepancies between the generated self-explanation and the model's decision process. In controlled settings, we show that this approach reduces the tendency of the LLM to generate misleading explanations.

Social Aspects · Privacy-preserving Statistics and Machine Learning

Ben Jacobsen, Kassem Fawaz

We study the classic problem of prediction with expert advice under the constraint of local differential privacy (LDP). In this context, we first show that a classical algorithm naturally satisfies LDP and then design two new algorithms that improve it: RW-AdaBatch and RW-Meta. For RW-AdaBatch, we exploit the limited-switching behavior induced by LDP to provide a novel form of privacy amplification that grows stronger on easier data, analogous to the shuffle model in offline learning. Drawing on the theory of random walks, we prove that this improvement carries essentially no utility cost. For RW-Meta, we develop a general method for privately selecting between experts that are themselves non-trivial learning algorithms, and we show that in the context of LDP this carries no extra privacy cost. In contrast, prior work has only considered data-independent experts. We also derive formal regret bounds that scale inversely with the degree of independence between experts. Our analysis is supplemented by evaluation on real-world data reported by hospitals during the COVID-19 pandemic; RW-Meta outperforms both the classical baseline and a state-of-the-art \textit{central} DP algorithm by 1.5-3$\times$ on the task of predicting which hospital will report the highest density of COVID patients each week.

Social Aspects · Fairness, Equity, Justice and Safety

Wonje Jeung, Sangyeon Yoon, Yoonjun Cho, Dongjae Jeon, Sangwoo Shin, Hyesoo Hong, Albert No

Diffusion large language models (dLLMs) enable any-order generation, but this flexibility enlarges the attack surface: harmful spans may appear at arbitrary positions, and template-based prefilling attacks such as DIJA bypass response-level refusals. We introduce A2D (Any-Order, Any-Step Defense), a token-level alignment method that aligns dLLMs to emit an [EOS] refusal signal whenever harmful content arises. By aligning safety directly at the token-level under randomized masking, A2D achieves robustness to both any-decoding-order and any-step prefilling attacks under various conditions. It also enables real-time monitoring: dLLMs may begin a response but automatically terminate if unsafe continuation emerges. On safety benchmarks, A2D consistently prevents the generation of harmful outputs, slashing DIJA success rates from over 80\% to near-zero (1.3\% on LLaDA-8B-Instruct, 0.0\% on Dream-v0-Instruct-7B), and thresholded [EOS] probabilities allow early rejection, yielding up to 19.3× faster safe termination.