Jailbreak attacks to Large audio-language models (LALMs) are studied recently, but they exclusively focused on the attack scenario where the adversary can fully manipulate user prompts (named strong adversary) and limited in effectiveness, applicability, and practicability. In this work, we first conduct an extensive evaluation showing that advanced text jailbreak attacks cannot be easily ported to end-to-end LALMs via text-to-speech (TTS) techniques. We then propose AUDIOJAILBREAK, a novel audio jailbreak attack, featuring (1) asynchrony: the jailbreak audios do not need to align with user prompts in the time axis by crafting suffixal jailbreak audios; (2) universality: a single jailbreak perturbation is effective for different prompts by incorporating multiple prompts into the perturbation generation; (3) stealthiness: the malicious intent of jailbreak audios is concealed by proposing various intent concealment strategies; and (4) over-the-air robustness: the jailbreak audios remain effective when being played over the air by incorporating reverberation into the perturbation generation. In contrast, all prior audio jailbreak attacks cannot offer asynchrony, universality, stealthiness, and/or over-the-air robustness. Moreover, AUDIOJAILBREAK is also applicable to a more practical and broader attack scenario where the adversary cannot fully manipulate user prompts (named weak adversary). Extensive experiments with thus far the most LALMs demonstrate the high effectiveness of AUDIOJAILBREAK, in particular, it can jailbreak openAI's GPT-4o-Audio and bypass Meta's Llama-Guard-3 safeguard, in the weak adversary scenario. We highlight that our work peeks into the security implications of audio jailbreak attacks against LALMs, and realistically fosters improving their robustness, especially for the newly proposed weak adversary.

Formal verification provides critical security assurances for neural networks, yet its practical application suffers from the long verification time. This work introduces a novel method for training verification-friendly neural networks, which are robust, easy to verify, and relatively accurate. Our method integrates neuron behavior consistency into the training process, making neuron activation states remain consistent across different inputs within a local neighborhood. This reduces the number of unstable neurons and tightens the bounds of neurons thereby enhancing the network's verifiability. We evaluated our method using the MNIST, Fashion-MNIST, and CIFAR-10 datasets with various network architectures. The experimental results demonstrate that networks trained using our method are verification-friendly across different radii and architectures, whereas other tools fail to maintain verifiability as the radius increases. Additionally, we show that our method can be combined with existing approaches to further improve the verifiability of networks.

As a new programming paradigm, deep learning (DL) has achieved impressive performance in areas such as image processing and speech recognition, and has expanded its application to solve many real-world problems. However, neural networks and DL are normally black-box systems; even worse, DL-based software are vulnerable to threats from abnormal examples, such as adversarial and backdoored examples constructed by attackers with malicious intentions as well as unintentionally mislabeled samples. Therefore, it is important and urgent to detect such abnormal examples. Although various detection approaches have been proposed respectively addressing some specific types of abnormal examples, they suffer from some limitations; until today, this problem is still of considerable interest. In this work, we first propose a novel characterization to distinguish abnormal examples from normal ones based on the observation that abnormal examples have significantly different (adversarial) robustness from normal ones. We systemically analyze those three different types of abnormal samples in terms of robustness and find that they have different characteristics from normal ones. As robustness measurement is computationally expensive and hence can be challenging to scale to large networks, we then propose to effectively and efficiently measure robustness of an input sample using the cost of adversarially attacking the input, which was originally proposed to test robustness of neural networks against adversarial examples. Next, we propose a novel detection method, named attack as detection (A 2 D for short), which uses the cost of adversarially attacking an input instead of robustness to check if it is abnormal. Our detection method is generic, and various adversarial attack methods could be leveraged. Extensive experiments show that A 2 D is more effective than recent promising approaches that were proposed to detect only one specific type of abnormal examples. We also thoroughly discuss possible adaptive attack methods to our adversarial example detection method and show that A 2 D is still effective in defending carefully designed adaptive adversarial attack methods—for example, the attack success rate drops to 0% on CIFAR10.

Speaker recognition systems (SRSs) have recently been shown to be vulnerable to adversarial attacks, raising significant security concerns. In this work, we systematically investigate transformation and adversarial training based defenses for securing SRSs. According to the characteristic of SRSs, we present 22 diverse transformations and thoroughly evaluate them using 7 recent promising adversarial attacks (4 white-box and 3 black-box) on speaker recognition. With careful regard for best practices in defense evaluations, we analyze the strength of transformations to withstand adaptive attacks. We also evaluate and understand their effectiveness against adaptive attacks when combined with adversarial training. Our study provides thirteen useful insights and findings, many of them are new or inconsistent with the conclusions in the image and speech recognition domains, e.g., variable and constant bit rate speech compressions have different performance, and some non-differentiable transformations remain effective against current promising evasion techniques which often work well in the image domain. We demonstrate that the proposed novel feature-level transformation combined with adversarial training is rather effective compared to the sole adversarial training in a complete white-box setting, e.g., increasing the accuracy by 13.62% and attack cost by two orders of magnitude, while other transformations do not necessarily improve the overall defense capability. This work sheds further light on the research directions in this field. We also release our evaluation platformSpeakerGuardto foster further research.

As a new programming paradigm, neural-network-based machine learning has expanded its application to many real-world problems. Due to the black-box nature of neural networks, verifying and explaining their behavior are becoming increasingly important, especially when they are deployed in safety-critical applications. Existing verification work mostly focuses on qualitative verification, which asks whether there exists an input (in a specified region) for a neural network such that a property (e.g., local robustness) is violated. However, in many practical applications, such an (adversarial) input almost surely exists, which makes a qualitative answer less meaningful. In this work, we study a more interesting yet more challenging problem, i.e.,quantitativeverification of neural networks, which asks how often a property is satisfied or violated. We target binarized neural networks (BNNs), the 1-bit quantization of general neural networks. BNNs have attracted increasing attention in deep learning recently, as they can drastically reduce memory storage and execution time with bit-wise operations, which is crucial in recourse-constrained scenarios, e.g., embedded devices for Internet of Things. Toward quantitative verification of BNNs, we propose a novel algorithmic approach for encoding BNNs as Binary Decision Diagrams (BDDs), a widely studied model in formal verification and knowledge representation. By exploiting the internal structure of the BNNs, our encoding translates the input-output relation of blocks in BNNs to cardinality constraints, which are then encoded by BDDs. Based on the new BDD encoding, we develop a quantitative verification framework for BNNs where precise and comprehensive analysis of BNNs can be performed. To improve the scalability of BDD encoding, we also investigate parallelization strategies at various levels. We demonstrate applications of our framework by providing quantitative robustness verification and interpretability for BNNs. An extensive experimental evaluation confirms the effectiveness and efficiency of our approach.

Deep learning has become a promising programming paradigm in software development, owing to its surprising performance in solving many challenging tasks. Deep neural networks (DNNs) are increasingly being deployed in practice, but are limited on resource-constrained devices owing to their demand for computational power. Quantization has emerged as a promising technique to reduce the size of DNNs with comparable accuracy as their floating-point numbered counterparts. The resulting quantized neural networks (QNNs) can be implemented energy-efficiently. Similar to their floating-point numbered counterparts, quality assurance techniques for QNNs, such as testing and formal verification, are essential but are currently less explored. In this work, we propose a novel and efficient formal verification approach for QNNs. In particular, we are the first to propose an encoding that reduces the verification problem of QNNs into the solving of integer linear constraints, which can be solved using off-the-shelf solvers. Our encoding is both sound and complete. We demonstrate the application of our approach on local robustness verification and maximum robustness radius computation. We implement our approach in a prototype tool QVIP and conduct a thorough evaluation. Experimental results on QNNs with different quantization bits confirm the effectiveness and efficiency of our approach, e.g., two orders of magnitude faster and able to solve more verification tasks in the same time limit than the state-of-the-art methods.

Neural network based (NN-based) classifiers are known vulnerable against adversarial examples, namely, adding slight perturbations to a benign image cause a classifier to make a false prediction. To evaluate the robustness of NN-based classifiers against adversarial examples, numerous adversarial attacks with high success rates have been proposed recently. NN-based image classifiers usually normalize valid images (e.g., RGB image where the value at each coordinate is an integer between 0 and 255) into a real continuous domain (e.g., 3-dimensional matrix where the value at each coordinate is a real number between 0 and 1) and make classification decisions on the normalized images. However, adversarial examples crafted in a real continuous domain may become benign once they are denormalized back into the corresponding discrete integer domain, known as the discretization problem. This problem has been mentioned in some prior works but received relatively limited attention. In this work, we report the first comprehensive study of existing works to understand the impacts of the discretization problem. By analyzing 35 representative methods and empirically studying 20 representative open source tools, we found 29/35 (theoretically) and 14/20 (empirically) are affected by the discretization problem, e.g., the success rate could dramatically drop from 100 to 10 percent after the domain transformation. As the first step towards addressing this problem in a black-box scenario, we propose a novel derivative-free optimization method, which can directly craft adversarial examples in the discrete integer domain. Experimental results show that the method achieves nearly 100 percent attack success rates for both targeted and untargeted attacks, comparable to the most popular white-box methods (FGSM, BIM and C&W), and significantly outperforms representative black-box methods (ZOO, AutoZOOM, NES-PGD, Bandits, FD, FD-PSO and GenAttack). Our results suggest that the discretization problem should be treated more seriously, and the discrete optimization algorithms show a promising future in crafting effective black-box attacks.

Abstract Verifying and explaining the behavior of neural networks is becoming increasingly important, especially when they are deployed in safety-critical applications. In this paper, we study verification and interpretability problems for Binarized Neural Networks (BNNs), the 1-bit quantization of general real-numbered neural networks. Our approach is to encode BNNs into Binary Decision Diagrams (BDDs), which is done by exploiting the internal structure of the BNNs. In particular, we translate the input-output relation of blocks in BNNs to cardinality constraints which are in turn encoded by BDDs. Based on the encoding, we develop a quantitative framework for BNNs where precise and comprehensive analysis of BNNs can be performed. We demonstrate the application of our framework by providing quantitative robustness analysis and interpretability for BNNs. We implement a prototype tool and carry out extensive experiments, confirming the effectiveness and efficiency of our approach.

As a new programming paradigm, deep learning has expanded its application to many real-world problems. At the same time, deep learning based software are found to be vulnerable to adversarial attacks. Though various defense mechanisms have been proposed to improve robustness of deep learning software, many of them are ineffective against adaptive attacks. In this work, we propose a novel characterization to distinguish adversarial examples from benign ones based on the observation that adversarial examples are significantly less robust than benign ones. As existing robustness measurement does not scale to large networks, we propose a novel defense framework, named attack as defense (A2D), to detect adversarial examples by effectively evaluating an example’s robustness. A2D uses the cost of attacking an input for robustness evaluation and identifies those less robust examples as adversarial since less robust examples are easier to attack. Extensive experiment results on MNIST, CIFAR10 and ImageNet show that A2D is more effective than recent promising approaches. We also evaluate our defense against potential adaptive attacks and show that A2D is effective in defending carefully designed adaptive attacks, e.g., the attack success rate drops to 0% on CIFAR10.

Speaker recognition (SR) is widely used in our daily life as a biometric authentication or identification mechanism. The popularity of SR brings in serious security concerns, as demonstrated by recent adversarial attacks. However, the impacts of such threats in the practical black-box setting are still open, since current attacks consider the white-box setting only.In this paper, we conduct the first comprehensive and systematic study of the adversarial attacks on SR systems (SRSs) to understand their security weakness in the practical black-box setting. For this purpose, we propose an adversarial attack, named FAKEBOB, to craft adversarial samples. Specifically, we formulate the adversarial sample generation as an optimization problem, incorporated with the confidence of adversarial samples and maximal distortion to balance between the strength and imperceptibility of adversarial voices. One key contribution is to propose a novel algorithm to estimate the score threshold, a feature in SRSs, and use it in the optimization problem to solve the optimization problem. We demonstrate that FAKEBOB achieves 99% targeted attack success rate on both open-source and commercial systems. We further demonstrate that FAKEBOB is also effective on both open-source and commercial systems when playing over the air in the physical world. Moreover, we have conducted a human study which reveals that it is hard for human to differentiate the speakers of the original and adversarial voices. Last but not least, we show that four promising defense methods for adversarial attack from the speech recognition domain become ineffective on SRSs against FAKEBOB, which calls for more effective defense methods. We highlight that our study peeks into the security implications of adversarial attacks on SRSs, and realistically fosters to improve the security robustness of SRSs.