Large Language Models (LLMs) are increasingly integrated with external tools. While these integrations can significantly improve the functionality of LLMs, they also create a new attack surface where confidential data may be disclosed between different components. Specifically, malicious tools can exploit vulnerabilities in the LLM itself to manipulate the model and compromise the data of other services, raising the question of how private data can be protected in the context of LLM integrations. In this work, we provide a systematic way of evaluating confidentiality in LLM-integrated systems. For this, we formalize a "secret key" game that can capture the ability of a model to conceal private information. This enables us to compare the vulnerability of a model against confidentiality attacks and also the effectiveness of different defense strategies. In this framework, we evaluate eight previously published attacks and four defenses. We find that current defenses lack generalization across attack strategies. Building on this analysis, we propose a method for robustness fine-tuning, inspired by adversarial training. This approach is effective in lowering the success rate of attackers and in improving the system's resilience against unknown attacks.
Evaluating the adversarial robustness of deep networks to gradient-based attacks is challenging. While most attacks consider $\ell_2$- and $\ell_\infty$-norm constraints to craft input perturbations, only a few investigate sparse $\ell_1$- and $\ell_0$-norm attacks. In particular, $\ell_0$-norm attacks remain the least studied due to the inherent complexity of optimizing over a non-convex and non-differentiable constraint. However, evaluating adversarial robustness under these attacks could reveal weaknesses otherwise left untested with more conventional $\ell_2$- and $\ell_\infty$-norm attacks. In this work, we propose a novel $\ell_0$-norm attack, called $\sigma$-zero, which leverages an ad hoc differentiable approximation of the $\ell_0$ norm to facilitate gradient-based optimization, and an adaptive projection operator to dynamically adjust the trade-off between loss minimization and perturbation sparsity. Extensive evaluations using MNIST, CIFAR10, and ImageNet datasets, involving robust and non-robust models, show that $\sigma$-zero finds minimum $\ell_0$-norm adversarial examples without requiring any time-consuming hyperparameter tuning, and that it outperforms all competing sparse attacks in terms of success rate, perturbation size, and scalability.
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.
There is a growing interest in using Large Language Models (LLMs) as agents to tackle real-world tasks that may require assessing complex situations. Yet, we have a limited understanding of LLMs' reasoning and decision-making capabilities, partly stemming from a lack of dedicated evaluation benchmarks. As negotiating and compromising are key aspects of our everyday communication and collaboration, we propose using scorable negotiation games as a new evaluation framework for LLMs. We create a testbed of diverse text-based, multi-agent, multi-issue, semantically rich negotiation games, with easily tunable difficulty. To solve the challenge, agents need to have strong arithmetic, inference, exploration, and planning capabilities, while seamlessly integrating them. Via a systematic zero-shot Chain-of-Thought prompting (CoT), we show that agents can negotiate and consistently reach successful deals. We quantify the performance with multiple metrics and observe a large gap between GPT-4 and earlier models. Importantly, we test the generalization to new games and setups. Finally, we show that these games can help evaluate other critical aspects, such as the interaction dynamics between agents in the presence of greedy and adversarial players.
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.
Recently, large language models for code generation have achieved breakthroughs in several programming language tasks. Their advances in competition-level programming problems have made them an emerging pillar in AI-assisted pair programming. Tools such as GitHub Copilot are already part of the daily programming workflow and are used by more than a million developers. The training data for these models is usually collected from open-source repositories (e.g., GitHub) that contain software faults and security vulnerabilities. This unsanitized training data can lead language models to learn these vulnerabilities and propagate them in the code generation procedure. Given the wide use of these models in the daily workflow of developers, it is crucial to study the security aspects of these models systematically. In this work, we propose the first approach to automatically finding security vulnerabilities in black-box code generation models. To achieve this, we propose a novel black-box inversion approach based on few-shot prompting. We evaluate the effectiveness of our approach by examining code generation models in the generation of high-risk security weaknesses. We show that our approach automatically and systematically finds 1000s of security vulnerabilities in various code generation models, including the commercial black-box model GitHub Copilot.
Deep generative modeling has the potential to cause significant harm to society. Recognizing this threat, a magnitude of research into detecting so-called "Deepfakes" has emerged. This research most often focuses on the image domain, while studies exploring generated audio signals have, so-far, been neglected. In this paper we make three key contributions to narrow this gap. First, we provide researchers with an introduction to common signal processing techniques used for analyzing audio signals. Second, we present a novel data set, for which we collected nine sample sets from five different network architectures, spanning two languages. Finally, we supply practitioners with two baseline models, adopted from the signal processing community, to facilitate further research in this area.
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, which 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.
In the past few years, we observed a wide adoption of practical systems that use Automatic Speech Recognition (ASR) systems to improve human-machine interaction. Modern ASR systems are based on neural networks and prior research demonstrated that these systems are susceptible to adversarial examples, i.e., malicious audio inputs that lead to misclassification by the victim's network during the system's run time. The research question if ASR systems are also vulnerable to data poisoning attacks is still unanswered. In such an attack, a manipulation happens during the training phase of the neural network: an adversary injects malicious inputs into the training set such that the neural network's integrity and performance are compromised. In this paper, we present the first data poisoning attack in the audio domain, called VENOMAVE. Prior work in the image domain demonstrated several types of data poisoning attacks, but they cannot be applied to the audio domain. The main challenge is that we need to attack a time series of inputs. To enforce a targeted misclassification in an ASR system, we need to carefully generate a specific sequence of disturbed inputs for the target utterance, which will eventually be decoded to the desired sequence of words. More specifically, the adversarial goal is to produce a series of misclassification tasks and in each of them, we need to poison the system to misrecognize each frame of the target file. To demonstrate the practical feasibility of our attack, we evaluate VENOMAVE on an ASR system that detects sequences of digits from 0 to 9. When poisoning only 0.94% of the dataset on average, we achieve an attack success rate of 83.33%. We conclude that data poisoning attacks against ASR systems represent a real threat that needs to be considered.
Machine learning systems and also, specifically, automatic speech recognition (ASR) systems are vulnerable against adversarial attacks, where an attacker maliciously changes the input. In the case of ASR systems, the most interesting cases are targeted attacks, in which an attacker aims to force the system into recognizing given target transcriptions in an arbitrary audio sample. The increasing number of sophisticated, quasi imperceptible attacks raises the question of countermeasures. In this paper, we focus on hybrid ASR systems and compare four acoustic models regarding their ability to indicate uncertainty under attack: a feed-forward neural network and three neural networks specifically designed for uncertainty quantification, namely a Bayesian neural network, Monte Carlo dropout, and a deep ensemble. We employ uncertainty measures of the acoustic model to construct a simple one-class classification model for assessing whether inputs are benign or adversarial. Based on this approach, we are able to detect adversarial examples with an area under the receiving operator curve score of more than 0.99. The neural networks for uncertainty quantification simultaneously diminish the vulnerability to the attack, which is reflected in a lower recognition accuracy of the malicious target text in comparison to a standard hybrid ASR system.