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.

Recent research shows that executable and linkable format (ELF) malware detection models based on deep learning are vulnerable to adversarial attacks. The most commonly used method in previous work is adversarial training to defend adversarial examples. Nevertheless, it is inefficient and only effective for specific adversarial attacks. Given that the perturbation byte insertion positions of existing adversarial malware generation methods are relatively fixed, we propose a new method to detect adversarial ELF malware. Using model interpretation techniques, we analyze the decision-making basis of the malware detection model and extract the features of adversarial examples. We further use anomaly detection techniques to identify adversarial examples. As an add-on module of the malware detection model, the proposed method does not require modifying the original model and does not need to retrain the model. Evaluating results show that the method can effectively defend the adversarial attacks against the malware detection model.

With the construction of smart cities, the number of Internet of Things (IoT) devices is growing rapidly, leading to an explosive growth of malware designed for IoT devices. These malware pose a serious threat to the security of IoT devices. The traditional malware classification methods mainly rely on feature engineering. To improve accuracy, a large number of different types of features will be extracted from malware files in these methods. That brings a high complexity to the classification. To solve these issues, a malware classification method based on Word2Vec and Multilayer Perception (MLP) is proposed in this article. First, for one malware sample, Word2Vec is used to calculate a word vector for all bytes of the binary file and all instructions in the assembly file. Second, we combine these vectors into a 256x256x2-dimensional matrix. Finally, we designed a deep learning network structure based on MLP to train the model. Then the model is used to classify the testing samples. The experimental results prove that the method has a high accuracy of 99.54%.

The traditional machine learning-based malware classification methods are mainly based on feature engineering. In order to improve accuracy, many features will be extracted from malware files in these methods. That brings a high complexity to the classification. To solve this issue, this paper proposes a malware classification method based on the word vector of bytes in the malware sample and Multilayer Perception (MLP). A malware sample consists of large number of bytes with values ranging from 0x00 to 0xFF. Therefore, every malware sample could be considered as a document written by bytes. And this document could be divided into sentences based on padding or meaningless bytes. In this paper, first, we use Word2Vec to calculate a 256 dimensions word vector for each byte. Second, we combine them into a matrix in ascending order. Third, we use MLP to train the model on the training samples. Finally, we use the trained model to classify the testing samples. The experimental results show that the method has a high accuracy of 98.89%.