Zeyu Mao

Jiongxiao Wang, Fangzhou Wu, Wendi Li et al.

Large language models (LLMs) have been widely deployed as the backbone with additional tools and text information for real-world applications. However, integrating external information into LLM-integrated applications raises significant security concerns. Among these, prompt injection attacks are particularly threatening, where malicious instructions injected in the external text information can exploit LLMs to generate answers as the attackers desire. While both training-time and test-time defense methods have been developed to mitigate such attacks, the unaffordable training costs associated with training-time methods and the limited effectiveness of existing test-time methods make them impractical. This paper introduces a novel test-time defense strategy, named Formatting AuThentication with Hash-based tags (FATH). Unlike existing approaches that prevent LLMs from answering additional instructions in external text, our method implements an authentication system, requiring LLMs to answer all received instructions with a security policy and selectively filter out responses to user instructions as the final output. To achieve this, we utilize hash-based authentication tags to label each response, facilitating accurate identification of responses according to the user's instructions and improving the robustness against adaptive attacks. Comprehensive experiments demonstrate that our defense method can effectively defend against indirect prompt injection attacks, achieving state-of-the-art performance under Llama3 and GPT3.5 models across various attack methods. Our code is released at: https://github.com/Jayfeather1024/FATH

Shuowei Jin, Francis Y. Yan, Cheng Tan et al.

The increasing adoption of neural networks in learning-augmented systems highlights the importance of model safety and robustness, particularly in safety-critical domains. Despite progress in the formal verification of neural networks, current practices require users to manually define model specifications -- properties that dictate expected model behavior in various scenarios. This manual process, however, is prone to human error, limited in scope, and time-consuming. In this paper, we introduce AutoSpec, the first framework to automatically generate comprehensive and accurate specifications for neural networks in learning-augmented systems. We also propose the first set of metrics for assessing the accuracy and coverage of model specifications, establishing a benchmark for future comparisons. Our evaluation across four distinct applications shows that AutoSpec outperforms human-defined specifications as well as two baseline approaches introduced in this study.

Patrick Wlazlo, Abhijeet Sahu, Zeyu Mao et al.

Man-in-The-Middle (MiTM) attacks present numerous threats to a smart grid. In a MiTM attack, an intruder embeds itself within a conversation between two devices to either eavesdrop or impersonate one of the devices, making it appear to be a normal exchange of information. Thus, the intruder can perform false data injection (FDI) and false command injection (FCI) attacks that can compromise power system operations, such as state estimation, economic dispatch, and automatic generation control (AGC). Very few researchers have focused on MiTM methods that are difficult to detect within a smart grid. To address this, we are designing and implementing multi-stage MiTM intrusions in an emulation-based cyber-physical power system testbed against a large-scale synthetic grid model to demonstrate how such attacks can cause physical contingencies such as misguided operation and false measurements. MiTM intrusions create FCI, FDI, and replay attacks in this synthetic power grid. This work enables stakeholders to defend against these stealthy attacks, and we present detection mechanisms that are developed using multiple alerts from intrusion detection systems and network monitoring tools. Our contribution will enable other smart grid security researchers and industry to develop further detection mechanisms for inconspicuous MiTM attacks.

Abhijeet Sahu, Zeyu Mao, Patrick Wlazlo et al.

Cyberattacks can cause a severe impact on power systems unless detected early. However, accurate and timely detection in critical infrastructure systems presents challenges, e.g., due to zero-day vulnerability exploitations and the cyber-physical nature of the system coupled with the need for high reliability and resilience of the physical system. Conventional rule-based and anomaly-based intrusion detection system (IDS) tools are insufficient for detecting zero-day cyber intrusions in the industrial control system (ICS) networks. Hence, in this work, we show that fusing information from multiple data sources can help identify cyber-induced incidents and reduce false positives. Specifically, we present how to recognize and address the barriers that can prevent the accurate use of multiple data sources for fusion-based detection. We perform multi-source data fusion for training IDS in a cyber-physical power system testbed where we collect cyber and physical side data from multiple sensors emulating real-world data sources that would be found in a utility and synthesizes these into features for algorithms to detect intrusions. Results are presented using the proposed data fusion application to infer False Data and Command injection-based Man-in- The-Middle (MiTM) attacks. Post collection, the data fusion application uses time-synchronized merge and extracts features followed by pre-processing such as imputation and encoding before training supervised, semi-supervised, and unsupervised learning models to evaluate the performance of the IDS. A major finding is the improvement of detection accuracy by fusion of features from cyber, security, and physical domains. Additionally, we observed the co-training technique performs at par with supervised learning methods when fed with our features.