Márk Jelasity

Levente Halmosi, Mark Jelasity

Achieving robustness against adversarial input perturbation is an important and intriguing problem in machine learning. In the area of semantic image segmentation, a number of adversarial training approaches have been proposed as a defense against adversarial perturbation, but the methodology of evaluating the robustness of the models is still lacking, compared to image classification. Here, we demonstrate that, just like in image classification, it is important to evaluate the models over several different and hard attacks. We propose a set of gradient based iterative attacks and show that it is essential to perform a large number of iterations. We include attacks against the internal representations of the models as well. We apply two types of attacks: maximizing the error with a bounded perturbation, and minimizing the perturbation for a given level of error. Using this set of attacks, we show for the first time that a number of models in previous work that are claimed to be robust are in fact not robust at all. We then evaluate simple adversarial training algorithms that produce reasonably robust models even under our set of strong attacks. Our results indicate that a key design decision to achieve any robustness is to use only adversarial examples during training. However, this introduces a trade-off between robustness and accuracy.

Zhixiang Guo, Siyuan Liang, Shi Fu et al.

Despite the growing use of world models as decision-making agents, their adversarial robustness remains underexplored due to the lack of dedicated automated evaluation methods. A key obstacle is that attack evaluation must be both accurate and efficient: weak manually tuned attacks can overestimate robustness, while exhaustive hyperparameter search is prohibitively expensive because each candidate requires closed-loop rollouts through learned latent dynamics. We introduce WMAttack, an automated attack-search framework for adversarial evaluation of world-model agents. WMAttack formulates robustness evaluation as a finite-budget search over attack configurations, including attack families, perturbation budgets, optimization steps, restarts, and allocation rules. To improve search accuracy, Self-Correcting Attack Search (SCAS) refines the attack proposal distribution using feedback from reward degradation, action instability, runtime cost, and rollout variability. To improve search efficiency, Representation-Guided Attack Retrieval (RGAR) retrieves effective historical configurations from representation-similar tasks, providing a warm start for unseen environments. We provide a theoretical explanation showing that proposal refinement improves finite-budget search when it shifts probability mass toward high-utility attacks. Across Atari and DeepMind Control tasks, WMAttack consistently discovers stronger attacks than the evaluated baselines, improving normalized reward drop from 0.497 to 1.034 on DreamerV3 Atari and from 0.319 to 0.682 on DMC. Ablations further show that RGAR improves initial candidate quality and SCAS improves final attack utility under fixed evaluation budgets.

Arpad Berta, Gabor Danner, Istvan Hegedus et al.

Detecting semantic backdoors in classification models--where some classes can be activated by certain natural, but out-of-distribution inputs--is an important problem that has received relatively little attention. Semantic backdoors are significantly harder to detect than backdoors that are based on trigger patterns due to the lack of such clearly identifiable patterns. We tackle this problem under the assumption that the clean training dataset and the training recipe of the model are both known. These assumptions are motivated by a consumer protection scenario, in which the responsible authority performs mystery shopping to test a machine learning service provider. In this scenario, the authority uses the provider's resources and tools to train a model on a given dataset and tests whether the provider included a backdoor. In our proposed approach, the authority creates a reference model pool by training a small number of clean and poisoned models using trusted infrastructure, and calibrates a model distance threshold to identify clean models. We propose and experimentally analyze a number of approaches to compute model distances and we also test a scenario where the provider performs an adaptive attack to avoid detection. The most reliable method is based on requesting adversarial training from the provider. The model distance is best measured using a set of input samples generated by inverting the models in such a way as to maximize the distance from clean samples. With these settings, our method can often completely separate clean and poisoned models, and it proves to be superior to state-of-the-art backdoor detectors as well.

Levente Halmosi, Bálint Mohos, Márk Jelasity

Machine learning models are vulnerable to tiny adversarial input perturbations optimized to cause a very large output error. To measure this vulnerability, we need reliable methods that can find such adversarial perturbations. For image classification models, evaluation methodologies have emerged that have stood the test of time. However, we argue that in the area of semantic segmentation, a good approximation of the sensitivity to adversarial perturbations requires significantly more effort than what is currently considered satisfactory. To support this claim, we re-evaluate a number of well-known robust segmentation models in an extensive empirical study. We propose new attacks and combine them with the strongest attacks available in the literature. We also analyze the sensitivity of the models in fine detail. The results indicate that most of the state-of-the-art models have a dramatically larger sensitivity to adversarial perturbations than previously reported. We also demonstrate a size-bias: small objects are often more easily attacked, even if the large objects are robust, a phenomenon not revealed by current evaluation metrics. Our results also demonstrate that a diverse set of strong attacks is necessary, because different models are often vulnerable to different attacks.

András Balogh, Márk Jelasity

Generative sequence models are typically trained on sample sequences from natural or formal languages. It is a crucial question whether -- or to what extent -- sample-based training is able to capture the true structure of these languages, often referred to as the ``world model''. Theoretical results indicate that we can hope for soundness at best, that is, generating valid sequences, but not necessarily all of them. However, it is still important to have practical tools that are able to verify whether a given sequence model is sound. In this study, we focus on chess, as it is a domain that provides enough complexity while having a simple rule-based world model. We propose adversarial sequence generation for verifying the soundness of the sequence model. Our adversaries generate valid sequences so as to force the sequence model to generate an invalid next move prediction. Apart from the falsification of soundness, this method is also suitable for a more fine-grained analysis of the failure modes and the effects of different choices during training. To demonstrate this, we propose a number of methods for adversarial sequence generation and evaluate the approach on a large set of chess models. We train models on random as well as high-quality chess games, using several training recipes. We find that none of the models are sound, but some training techniques and dataset choices are able to improve soundness remarkably. We also investigate the potential application of board state probes in both our training and attack methods. Our findings indicate that the extracted board states have no causal role in next token prediction in most of the models.

Attila Szász, Balázs Bánhelyi, Márk Jelasity

The ultimate goal of verification is to guarantee the safety of deployed neural networks. Here, we claim that all the state-of-the-art verifiers we are aware of fail to reach this goal. Our key insight is that theoretical soundness (bounding the full-precision output while computing with floating point) does not imply practical soundness (bounding the floating point output in a potentially stochastic environment). We prove this observation for the approaches that are currently used to achieve provable theoretical soundness, such as interval analysis and its variants. We also argue that achieving practical soundness is significantly harder computationally. We support our claims empirically as well by evaluating several well-known verification methods. To mislead the verifiers, we create adversarial networks that detect and exploit features of the deployment environment, such as the order and precision of floating point operations. We demonstrate that all the tested verifiers are vulnerable to our new deployment-specific attacks, which proves that they are not practically sound.

Zhixiang Guo, Siyuan Liang, Andras Balogh et al.

Generative world models (WMs) are increasingly used to synthesize controllable, sensor-conditioned driving videos, yet their reliance on physical priors exposes novel attack surfaces. In this paper, we present Physical-Conditioned World Model Attack (PhysCond-WMA), the first white-box world model attack that perturbs physical-condition channels, such as HDMap embeddings and 3D-box features, to induce semantic, logic, or decision-level distortion while preserving perceptual fidelity. PhysCond-WMA is optimized in two stages: (1) a quality-preserving guidance stage that constrains reverse-diffusion loss below a calibrated threshold, and (2) a momentum-guided denoising stage that accumulates target-aligned gradients along the denoising trajectory for stable, temporally coherent semantic shifts. Extensive experimental results demonstrate that our approach remains effective while increasing FID by about 9% on average and FVD by about 3.9% on average. Under the targeted attack setting, the attack success rate (ASR) reaches 0.55. Downstream studies further show tangible risk, which using attacked videos for training decreases 3D detection performance by about 4%, and worsens open-loop planning performance by about 20%. These findings has for the first time revealed and quantified security vulnerabilities in generative world models, driving more comprehensive security checkers.

András Balogh, Márk Jelasity

Measuring the similarity of the internal representations of deep neural networks is an important and challenging problem. Model stitching has been proposed as a possible approach, where two half-networks are connected by mapping the output of the first half-network to the input of the second one. The representations are considered functionally similar if the resulting stitched network achieves good task-specific performance. The mapping is normally created by training an affine stitching layer on the task at hand while freezing the two half-networks, a method called task loss matching. Here, we argue that task loss matching may be very misleading as a similarity index. For example, it can indicate very high similarity between very distant layers, whose representations are known to have different functional properties. Moreover, it can indicate very distant layers to be more similar than architecturally corresponding layers. Even more surprisingly, when comparing layers within the same network, task loss matching often indicates that some layers are more similar to a layer than itself. We argue that the main reason behind these problems is that task loss matching tends to create out-of-distribution representations to improve task-specific performance. We demonstrate that direct matching (when the mapping minimizes the distance between the stitched representations) does not suffer from these problems. We compare task loss matching, direct matching, and well-known similarity indices such as CCA and CKA. We conclude that direct matching strikes a good balance between the structural and functional requirements for a good similarity index.

Lea Hergert, Gábor Berend, Mario Szegedy et al.

Large language models (LLMs) achieve superhuman performance on complex reasoning tasks, yet often fail on much simpler problems, raising concerns about their reliability and interpretability. We investigate this paradox through a focused study with two key design features: simplicity, to expose basic failure modes, and scale, to enable comprehensive controlled experiments. We focus on set membership queries -- among the most fundamental forms of reasoning -- using tasks like ``Is apple an element of the set \{pear, plum, apple, raspberry\}?''. We conduct a systematic empirical evaluation across prompt phrasing, semantic structure, element ordering, and model choice. Our large-scale analysis reveals that LLM performance on this elementary task is consistently brittle, and unpredictable across all dimensions, suggesting that the models' ``understanding'' of the set concept is fragmented and convoluted at best. Our work demonstrates that the large-scale experiments enabled by the simplicity of the problem allow us to map and analyze the failure modes comprehensively, making this approach a valuable methodology for LLM evaluation in general.