Thorsten Eisenhofer

Thorsten Eisenhofer, Erwin Quiring, Jonas Möller et al.

The number of papers submitted to academic conferences is steadily rising in many scientific disciplines. To handle this growth, systems for automatic paper-reviewer assignments are increasingly used during the reviewing process. These systems use statistical topic models to characterize the content of submissions and automate the assignment to reviewers. In this paper, we show that this automation can be manipulated using adversarial learning. We propose an attack that adapts a given paper so that it misleads the assignment and selects its own reviewers. Our attack is based on a novel optimization strategy that alternates between the feature space and problem space to realize unobtrusive changes to the paper. To evaluate the feasibility of our attack, we simulate the paper-reviewer assignment of an actual security conference (IEEE S&P) with 165 reviewers on the program committee. Our results show that we can successfully select and remove reviewers without access to the assignment system. Moreover, we demonstrate that the manipulated papers remain plausible and are often indistinguishable from benign submissions.

Thorsten Eisenhofer, Doreen Riepel, Varun Chandrasekaran et al.

Machine unlearning aims to remove points from the training dataset of a machine learning model after training: e.g., when a user requests their data to be deleted. While many unlearning methods have been proposed, none of them enable users to audit the procedure. Furthermore, recent work shows a user is unable to verify whether their data was unlearnt from an inspection of the model parameter alone. Rather than reasoning about parameters, we propose to view verifiable unlearning as a security problem. To this end, we present the first cryptographic definition of verifiable unlearning to formally capture the guarantees of an unlearning system. In this framework, the server first computes a proof that the model was trained on a dataset D. Given a user's data point d requested to be deleted, the server updates the model using an unlearning algorithm. It then provides a proof of the correct execution of unlearning and that d is not part of D', where D' is the new training dataset (i.e., d has been removed). Our framework is generally applicable to different unlearning techniques that we abstract as admissible functions. We instantiate a protocol in the framework, based on cryptographic assumptions, using SNARKs and hash chains. Finally, we implement the protocol for three different unlearning techniques and validate its feasibility for linear regression, logistic regression, and neural networks.

Roei Schuster, Jin Peng Zhou, Thorsten Eisenhofer et al.

A learned database system uses machine learning (ML) internally to improve performance. We can expect such systems to be vulnerable to some adversarial-ML attacks. Often, the learned component is shared between mutually-distrusting users or processes, much like microarchitectural resources such as caches, potentially giving rise to highly-realistic attacker models. However, compared to attacks on other ML-based systems, attackers face a level of indirection as they cannot interact directly with the learned model. Additionally, the difference between the attack surface of learned and non-learned versions of the same system is often subtle. These factors obfuscate the de-facto risks that the incorporation of ML carries. We analyze the root causes of potentially-increased attack surface in learned database systems and develop a framework for identifying vulnerabilities that stem from the use of ML. We apply our framework to a broad set of learned components currently being explored in the database community. To empirically validate the vulnerabilities surfaced by our framework, we choose 3 of them and implement and evaluate exploits against these. We show that the use of ML cause leakage of past queries in a database, enable a poisoning attack that causes exponential memory blowup in an index structure and crashes it in seconds, and enable index users to snoop on each others' key distributions by timing queries over their own keys. We find that adversarial ML is an universal threat against learned components in database systems, point to open research gaps in our understanding of learned-systems security, and conclude by discussing mitigations, while noting that data leakage is inherent in systems whose learned component is shared between multiple parties.

David Pape, Sina Mavali, Thorsten Eisenhofer et al.

System prompts that include detailed instructions to describe the task performed by the underlying LLM can easily transform foundation models into tools and services with minimal overhead. They are often considered intellectual property, similar to the code of a software product, because of their crucial impact on the utility. However, extracting system prompts is easily possible. As of today, there is no effective countermeasure to prevent the stealing of system prompts, and all safeguarding efforts could be evaded. In this work, we propose an alternative to conventional system prompts. We introduce prompt obfuscation to prevent the extraction of the system prompt with little overhead. The core idea is to find a representation of the original system prompt that leads to the same functionality, while the obfuscated system prompt does not contain any information that allows conclusions to be drawn about the original system prompt. We evaluate our approach by comparing our obfuscated prompt output with the output of the original prompt, using eight distinct metrics to measure the lexical, character-level, and semantic similarity. We show that the obfuscated version is constantly on par with the original one. We further perform three different deobfuscation attacks with varying attacker knowledge--covering both black-box and white-box conditions--and show that in realistic attack scenarios an attacker is unable to extract meaningful information. Overall, we demonstrate that prompt obfuscation is an effective mechanism to safeguard the intellectual property of a system prompt while maintaining the same utility as the original prompt.

Sina Mavali, David Pape, Jonathan Evertz et al.

Terminal agents are increasingly capable of executing complex, long-horizon tasks autonomously from a single user prompt. To do so, they must interpret instructions encountered in the environment (e.g., README files, code comments, stack traces) and determine their relevance to the task. This creates a fundamental challenge: relevant cues must be followed to complete a task, whereas irrelevant or misleading ones must be ignored. Existing benchmarks do not capture this ability. An agent may appear capable by blindly following all instructions, or appear robust by ignoring them altogether. We introduce TAB (Task Alignment Benchmark), a suite of 89 terminal tasks derived from Terminal-Bench 2.1. Each task is intentionally underspecified, with missing information provided as a necessary cue embedded in a natural environmental artifact, alongside a plausible but irrelevant distractor. Solving these tasks requires selectively using the cue while ignoring the distractor. Applying TAB to ten frontier agents reveals a systematic gap between task capability and task alignment. The strongest Terminal-Bench agent achieves high task completion but low task alignment on TAB. Evaluating six prompt-injection defenses further shows that suppressing distractor execution also suppresses the cues required for task completion. These results demonstrate that task-aligned agents require selective use of environmental instructions rather than blanket acceptance or rejection.

Jonas Möller, Erik Imgrund, Thorsten Eisenhofer et al.

Machine learning models are routinely deployed on a wide range of computing hardware. Although such hardware is typically expected to produce identical results, differences in its design can lead to small numerical variations during inference. In this work, we show that these variations can be exploited to create backdoors in machine learning models. The core idea is to shape the model's decision function such that it yields different predictions for the same input when executed on different hardware. This effect is achieved by locally moving the decision boundary close to a target input and then refining numerical deviations to flip the prediction on selected hardware. We empirically demonstrate that these hardware-triggered backdoors can be created reliably across common GPU accelerators. Our findings reveal a novel attack vector affecting the use of third-party models, and we investigate different defenses to counter this threat.

Erik Imgrund, Thorsten Eisenhofer, Konrad Rieck

AI-based systems, such as Google's GenCast, have recently redefined the state of the art in weather forecasting, offering more accurate and timely predictions of both everyday weather and extreme events. While these systems are on the verge of replacing traditional meteorological methods, they also introduce new vulnerabilities into the forecasting process. In this paper, we investigate this threat and present a novel attack on autoregressive diffusion models, such as those used in GenCast, capable of manipulating weather forecasts and fabricating extreme events, including hurricanes, heat waves, and intense rainfall. The attack introduces subtle perturbations into weather observations that are statistically indistinguishable from natural noise and change less than 0.1% of the measurements - comparable to tampering with data from a single meteorological satellite. As modern forecasting integrates data from nearly a hundred satellites and many other sources operated by different countries, our findings highlight a critical security risk with the potential to cause large-scale disruptions and undermine public trust in weather prediction.

Felix Weissberg, Lukas Pirch, Erik Imgrund et al.

Large language models (LLMs) excel in many tasks of software engineering, yet progress in leveraging them for vulnerability discovery has stalled in recent years. To understand this phenomenon, we investigate LLMs through the lens of classic code metrics. Surprisingly, we find that a classifier trained solely on these metrics performs on par with state-of-the-art LLMs for vulnerability discovery. A root-cause analysis reveals a strong correlation and a causal effect between LLMs and code metrics: When the value of a metric is changed, LLM predictions tend to shift by a corresponding magnitude. This dependency suggests that LLMs operate at a similarly shallow level as code metrics, limiting their ability to grasp complex patterns and fully realize their potential in vulnerability discovery. Based on these findings, we derive recommendations on how research should more effectively address this challenge.

Jonathan Evertz, Merlin Chlosta, Lea Schönherr et al.

The interaction between users and applications is increasingly shifted toward natural language by deploying Large Language Models (LLMs) as the core interface. The capabilities of these so-called agents become more capable the more tools and services they serve as an interface for, ultimately leading to agentic systems. Agentic systems use LLM-based agents as interfaces for most user interactions and various integrations with external tools and services. While these interfaces can significantly enhance the capabilities of the agentic system, they also introduce a new attack surface. Manipulated integrations, for example, can exploit the internal LLM and compromise sensitive data accessed through other interfaces. While previous work primarily focused on attacks targeting a model's alignment or the leakage of training data, the security of data that is only available during inference has escaped scrutiny so far. In this work, we demonstrate how the integration of LLMs into systems with external tool integration poses a risk similar to established prompt-based attacks, able to compromise the confidentiality of the entire system. Introducing a systematic approach to evaluate these confidentiality risks, we identify two specific attack scenarios unique to these agentic systems and formalize these into a tool-robustness framework designed to measure a model's ability to protect sensitive information. Our analysis reveals significant vulnerabilities across all tested models, highlighting an increased risk when models are combined with external tools.

Joel Frank, Franziska Herbert, Jonas Ricker et al.

AI-generated media has become a threat to our digital society as we know it. These forgeries can be created automatically and on a large scale based on publicly available technology. Recognizing this challenge, academics and practitioners have proposed a multitude of automatic detection strategies to detect such artificial media. However, in contrast to these technical advances, the human perception of generated media has not been thoroughly studied yet. In this paper, we aim at closing this research gap. We perform the first comprehensive survey into people's ability to detect generated media, spanning three countries (USA, Germany, and China) with 3,002 participants across audio, image, and text media. Our results indicate that state-of-the-art forgeries are almost indistinguishable from "real" media, with the majority of participants simply guessing when asked to rate them as human- or machine-generated. In addition, AI-generated media receive is voted more human like across all media types and all countries. To further understand which factors influence people's ability to detect generated media, we include personal variables, chosen based on a literature review in the domains of deepfake and fake news research. In a regression analysis, we found that generalized trust, cognitive reflection, and self-reported familiarity with deepfakes significantly influence participant's decision across all media categories.

David Pape, Sina Däubener, Thorsten Eisenhofer et al.

Model stealing aims at inferring a victim model's functionality at a fraction of the original training cost. While the goal is clear, in practice the model's architecture, weight dimension, and original training data can not be determined exactly, leading to mutual uncertainty during stealing. In this work, we explicitly tackle this uncertainty by generating multiple possible networks and combining their predictions to improve the quality of the stolen model. For this, we compare five popular uncertainty quantification models in a model stealing task. Surprisingly, our results indicate that the considered models only lead to marginal improvements in terms of label agreement (i.e., fidelity) to the stolen model. To find the cause of this, we inspect the diversity of the model's prediction by looking at the prediction variance as a function of training iterations. We realize that during training, the models tend to have similar predictions, indicating that the network diversity we wanted to leverage using uncertainty quantification models is not (high) enough for improvements on the model stealing task.

Thorsten Eisenhofer, Lea Schönherr, Joel Frank et al.

Adversarial examples seem to be inevitable. These specifically crafted inputs allow attackers to arbitrarily manipulate machine learning systems. Even worse, they often seem harmless to human observers. In our digital society, this poses a significant threat. For example, Automatic Speech Recognition (ASR) systems, which serve as hands-free interfaces to many kinds of systems, can be attacked with inputs incomprehensible for human listeners. The research community has unsuccessfully tried several approaches to tackle this problem. In this paper we propose a different perspective: We accept the presence of adversarial examples against ASR systems, but we require them to be perceivable by human listeners. By applying the principles of psychoacoustics, we can remove semantically irrelevant information from the ASR input and train a model that resembles human perception more closely. We implement our idea in a tool named DOMPTEUR and demonstrate that our augmented system, in contrast to an unmodified baseline, successfully focuses on perceptible ranges of the input signal. This change forces adversarial examples into the audible range, while using minimal computational overhead and preserving benign performance. To evaluate our approach, we construct an adaptive attacker that actively tries to avoid our augmentations and demonstrate that adversarial examples from this attacker remain clearly perceivable. Finally, we substantiate our claims by performing a hearing test with crowd-sourced human listeners.

Hojjat Aghakhani, Lea Schönherr, Thorsten Eisenhofer et al.

Despite remarkable improvements, automatic speech recognition is susceptible to adversarial perturbations. Compared to standard machine learning architectures, these attacks are significantly more challenging, especially since the inputs to a speech recognition system are time series that contain both acoustic and linguistic properties of speech. Extracting all recognition-relevant information requires more complex pipelines and an ensemble of specialized components. Consequently, an attacker needs to consider the entire pipeline. In this paper, we present VENOMAVE, the first training-time poisoning attack against speech recognition. Similar to the predominantly studied evasion attacks, we pursue the same goal: leading the system to an incorrect and attacker-chosen transcription of a target audio waveform. In contrast to evasion attacks, however, we assume that the attacker can only manipulate a small part of the training data without altering the target audio waveform at runtime. We evaluate our attack on two datasets: TIDIGITS and Speech Commands. When poisoning less than 0.17% of the dataset, VENOMAVE achieves attack success rates of more than 80.0%, without access to the victim's network architecture or hyperparameters. In a more realistic scenario, when the target audio waveform is played over the air in different rooms, VENOMAVE maintains a success rate of up to 73.3%. Finally, VENOMAVE achieves an attack transferability rate of 36.4% between two different model architectures.

Lea Schönherr, Maximilian Golla, Thorsten Eisenhofer et al.

Voice assistants like Amazon's Alexa, Google's Assistant, or Apple's Siri, have become the primary (voice) interface in smart speakers that can be found in millions of households. For privacy reasons, these speakers analyze every sound in their environment for their respective wake word like ''Alexa'' or ''Hey Siri,'' before uploading the audio stream to the cloud for further processing. Previous work reported on the inaccurate wake word detection, which can be tricked using similar words or sounds like ''cocaine noodles'' instead of ''OK Google.'' In this paper, we perform a comprehensive analysis of such accidental triggers, i.,e., sounds that should not have triggered the voice assistant, but did. More specifically, we automate the process of finding accidental triggers and measure their prevalence across 11 smart speakers from 8 different manufacturers using everyday media such as TV shows, news, and other kinds of audio datasets. To systematically detect accidental triggers, we describe a method to artificially craft such triggers using a pronouncing dictionary and a weighted, phone-based Levenshtein distance. In total, we have found hundreds of accidental triggers. Moreover, we explore potential gender and language biases and analyze the reproducibility. Finally, we discuss the resulting privacy implications of accidental triggers and explore countermeasures to reduce and limit their impact on users' privacy. To foster additional research on these sounds that mislead machine learning models, we publish a dataset of more than 1000 verified triggers as a research artifact.

Joel Frank, Thorsten Eisenhofer, Lea Schönherr et al.

Deep neural networks can generate images that are astonishingly realistic, so much so that it is often hard for humans to distinguish them from actual photos. These achievements have been largely made possible by Generative Adversarial Networks (GANs). While deep fake images have been thoroughly investigated in the image domain - a classical approach from the area of image forensics - an analysis in the frequency domain has been missing so far. In this paper, we address this shortcoming and our results reveal that in frequency space, GAN-generated images exhibit severe artifacts that can be easily identified. We perform a comprehensive analysis, showing that these artifacts are consistent across different neural network architectures, data sets, and resolutions. In a further investigation, we demonstrate that these artifacts are caused by upsampling operations found in all current GAN architectures, indicating a structural and fundamental problem in the way images are generated via GANs. Based on this analysis, we demonstrate how the frequency representation can be used to identify deep fake images in an automated way, surpassing state-of-the-art methods.

Lea Schönherr, Thorsten Eisenhofer, Steffen Zeiler et al.

Automatic speech recognition (ASR) systems can be fooled via targeted adversarial examples, which induce the ASR to produce arbitrary transcriptions in response to altered audio signals. However, state-of-the-art adversarial examples typically have to be fed into the ASR system directly, and are not successful when played in a room. The few published over-the-air adversarial examples fall into one of three categories: they are either handcrafted examples, they are so conspicuous that human listeners can easily recognize the target transcription once they are alerted to its content, or they require precise information about the room where the attack takes place, and are hence not transferable to other rooms. In this paper, we demonstrate the first algorithm that produces generic adversarial examples, which remain robust in an over-the-air attack that is not adapted to the specific environment. Hence, no prior knowledge of the room characteristics is required. Instead, we use room impulse responses (RIRs) to compute robust adversarial examples for arbitrary room characteristics and employ the ASR system Kaldi to demonstrate the attack. Further, our algorithm can utilize psychoacoustic methods to hide changes of the original audio signal below the human thresholds of hearing. In practical experiments, we show that the adversarial examples work for varying room setups, and that no direct line-of-sight between speaker and microphone is necessary. As a result, an attacker can create inconspicuous adversarial examples for any target transcription and apply these to arbitrary room setups without any prior knowledge.