With the advent of text-to-image models and concerns about their misuse, developers are increasingly relying on image safety classifiers to moderate their generated unsafe images. Yet, the performance of current image safety classifiers remains unknown for both real-world and AI-generated images. In this work, we propose UnsafeBench, a benchmarking framework that evaluates the effectiveness and robustness of image safety classifiers, with a particular focus on the impact of AI-generated images on their performance. First, we curate a large dataset of 10K real-world and AI-generated images that are annotated as safe or unsafe based on a set of 11 unsafe categories of images (sexual, violent, hateful, etc.). Then, we evaluate the effectiveness and robustness of five popular image safety classifiers, as well as three classifiers that are powered by general-purpose visual language models. Our assessment indicates that existing image safety classifiers are not comprehensive and effective enough to mitigate the multifaceted problem of unsafe images. Also, there exists a distribution shift between real-world and AI-generated images in image qualities, styles, and layouts, leading to degraded effectiveness and robustness. Motivated by these findings, we build a comprehensive image moderation tool called PerspectiveVision, which improves the effectiveness and robustness of existing classifiers, especially on AI-generated images. UnsafeBench and PerspectiveVision can aid the research community in better understanding the landscape of image safety classification in the era of generative AI.

Text-to-image models, such as Stable Diffusion (SD), undergo iterative updates to improve image quality and address concerns such as safety. Improvements in image quality are straightforward to assess. However, how model updates resolve existing concerns and whether they raise new questions remain unexplored. This study takes an initial step in investigating the evolution of text-to-image models from the perspectives of safety, bias, and authenticity. Our findings, centered on Stable Diffusion, indicate that model updates paint a mixed picture. While updates progressively reduce the generation of unsafe images, the bias issue, particularly in gender, intensifies. We also find that negative stereotypes either persist within the same Non-White race group or shift towards other Non-White race groups through SD updates, yet with minimal association of these traits with the White race group. Additionally, our evaluation reveals a new concern stemming from SD updates: State-of-the-art fake image detectors, initially trained for earlier SD versions, struggle to identify fake images generated by updated versions. We show that fine-tuning these detectors on fake images generated by updated versions achieves at least 96.6% accuracy across various SD versions, addressing this issue. Our insights highlight the importance of continued efforts to mitigate biases and vulnerabilities in evolving text-to-image models.

Effective utilization of large language models (LLMs), such as ChatGPT, relies on the quality of input prompts.This paper explores prompt engineering, specifically focusing on the disparity between experimentally designed prompts and real-world "in-the-wild" prompts.We analyze 10,538 in-the-wild prompts collected from various platforms and develop a framework that decomposes the prompts into eight key components.Our analysis shows that Role and Requirement are the most prevalent two components.Roles specified in the prompts, along with their capabilities, have become increasingly varied over time, signifying a broader range of application scenarios for LLMs.However, from the response of GPT-4, there is a marginal improvement with a specified role, whereas leveraging less prevalent components such as Capability and Demonstration can result in a more satisfying response.Overall, our work sheds light on the essential components of in-the-wild prompts and the effectiveness of these components on the broader landscape of LLM prompt engineering, providing valuable guidelines for the LLM community to optimize high-quality prompts.

A graph neural network (GNN) is a type of neural network that is specifically designed to process graph-structured data. Typically, GNNs can be implemented in two settings, including the transductive setting and the inductive setting. In the transductive setting, the trained model can only predict the labels of nodes that were observed at the training time. In the inductive setting, the trained model can be generalized to new nodes/graphs. Due to its flexibility, the inductive setting is the most popular GNN setting at the moment. Previous work has shown that transductive GNNs are vulnerable to a series of privacy attacks. However, a comprehensive privacy analysis of inductive GNN models is still missing. This paper fills the gap by conducting a systematic privacy analysis of inductive GNNs through the lens of link stealing attacks. We propose two types of link stealing attacks, i.e., posterior-only attacks and combined attacks. We define threat models of the posterior-only attacks with respect to node topology and the combined attacks by considering combinations of posteriors, node attributes, and graph features. Extensive evaluation on six real-world datasets demonstrates that inductive GNNs leak rich information that enables link stealing attacks with advantageous properties. Even attacks with no knowledge about graph structures can be effective. We also show that our attacks are robust to different node similarities and different graph features. As a counterpart, we investigate two possible defenses and discover they are ineffective against our attacks, which calls for more effective defenses.

As the defense shifts from vulnerability-centric to threat-centric and efficient security architecture can exclusively be constructed with adequate comprehension of the threat of the critical assets. In order to classify and identify the assets, the recognition and naming of the Web applications are the fundamental approaches. At present, the traditional Web application identification methods mainly rely on rules matching, which are extracted from the Web pages by manual analysis. This low coverage and labor-consuming method, which is not suitable for this time of explosive growth in Web applications and inevitably leaves some uncommon applications unrecognized and at risk. In this paper, we propose WAIN, an automatic method for Web application identification and naming, it first clusters different types of applications in numerous samples using K-Means algorithm, and then leverages a novel TF-IDF calculation method to extract keyword. After that, LDA is applied to explain why some parts of data are similar and extract possible fingerprints. Finally, WAIN utilizes filters and a statistic means to generate possible names for clusters. When evaluating, data from 30,000 instances of eight kinds of Web applications is processed, and the generated fingerprints and names can distinguish each type of application in the dataset. We manually checked all the results and found that fingerprints and at least one name that summarizes at least one of the product names, manufacturers, and functions are successfully generated for each kind of application.

As Advanced Persistent Threat (APT) becomes increasingly frequent around the world, security experts are starting to look at how to observe, predict, and mitigate the damage from APT attacks. In the meantime, the Internet of things devices are also risky and heavily exposed to the Internet, making them more easily used by hacker organizations to launch APT attacks. An excellent attacker can take down millions of Internet of things devices in a short time. Once the IoT botnet is built, attackers can use it to launch complex attacks which could damage Internet infrastructure and cause network disconnection. This paper proposes GroupTracer, a framework for observing and predicting the Internet of things attacks. GroupTracer is designed to automatically extract the TTP profiles (i.e., tactics, techniques, and procedures) that can describe the behavior of attackers through their tactics, techniques, and processes and dig out the potential attacker groups behind complex attacks. Firstly, it captures attacks by IoT honeypots and extracts relevant fields from logs. Then, attack behaviors are automatically mapped to the ATT&CK framework to achieve automatic TTP profiles extraction. After that, GroupTracer presents four feature groups, including TTP profiles, Time, IP, and URL features, a total of 18 features, mines potential attack groups through hierarchical clustering algorithm, and compares the clustering results with two baseline algorithms. As the ground truth labels are unknown, we apply three internal validation indexes to evaluate the cluster quantity. Experimental results showed that the proposed framework has achieved an excellent performance in exploiting potential groups as the Calinski–Harabasz index reaches 3416.93. Eventually, attack trees are generated for each cluster where nodes indicate attack commands and edges represent command sequences. These attack trees could help better understand each attack group’s actions and techniques.

Attackers upload webshell into a web server to achieve the purpose of stealing data, launching a DDoS attack, modifying files with malicious intentions, etc. Once these objects are accomplished, it will bring huge losses to website managers. With the gradual development of encryption and confusion technology, the most common detection approach using taint analysis and feature matching might become less useful. Instead of applying source file codes, POST contents, or all received traffic, this paper demonstrated an intelligent and efficient framework that employs precise sessions derived from the web logs to detect webshell communication. Features were extracted from the raw sequence data in web logs while a statistical method based on time interval was proposed to identify sessions specifically. Besides, the paper leveraged long short-term memory and hidden Markov model to constitute the framework, respectively. Finally, the framework was evaluated with real data. The experiment shows that the LSTM-based model can achieve a higher accuracy rate of 95.97% with a recall rate of 96.15%, which has a much better performance than the HMM-based model. Moreover, the experiment demonstrated the high efficiency of the proposed approach in terms of the quick detection without source code, especially when it only considers detecting for a period of time, as it takes 98.5% less time than the cited related approach to get the result. As long as the webshell behavior is detected, we can pinpoint the anomaly session and utilize the statistical method to find the webshell file accurately.