Qi Li

Smart contracts are foundational to DeFi, yet detecting business-logic vulnerabilities, i.e., execution is syntactically correct but violates a contract’s intended behavior, remains difficult. Most recent LLM-based approaches still follow a largely static, source-code–centric auditing paradigm. In practice, however, smart contracts are inherently stateful: many logic flaws emerge only under specific contract states and across particular state transitions. As a result, code-only auditing can produce false positives and miss true vulnerabilities. To address this gap, we present RcAudit, a system for detecting smart-contract logic vulnerabilities via runtime-context–enhanced LLM reasoning. RcAudit constructs a Contract Logic Knowledge Base that captures contract behavior from multiple perspectives to support LLM analysis. It then applies a state- and intent-aware fuzzing strategy to generate logic-critical runtime contexts, and performs LLM reasoning over selected contexts that exhibit potential intent–behavior inconsistencies to pinpoint vulnerabilities. We evaluate RcAudit on three real-world datasets. RcAudit identifies 74 of 132 known logic vulnerabilities with high precision (96.9%), achieving 117.6% higher detection than the strongest baseline. Importantly, RcAudit further uncovers 22 zero-day vulnerabilities in real-world projects.

Website Fingerprinting (WF) attacks exploit patterns in encrypted traffic to infer the websites visited by users, posing a serious threat to anonymous communication systems. Although recent WF techniques achieve over 0.9 accuracy in controlled experimental settings, most studies remain confined to single scenarios, overlooking the complexity of real-world environments. This paper presents the first systematic and comprehensive evaluation of existing WF attacks under diverse realistic conditions, including defense mechanisms, traffic drift, multi-tab browsing, early-stage detection, open-world settings, and few-shot scenarios. Experimental results show that many WF techniques with strong performance in isolated settings degrade significantly when facing other conditions. Since real-world environments often combine multiple challenges, current WF attacks are difficult to apply directly in practice. This study highlights the limitations of WF attacks and introduces a multidimensional evaluation framework, offering critical insights for developing more robust and practical WF attacks.

The widespread adoption and significant development costs associated with Graph Neural Networks (GNNs) increase vulnerability to model stealing attacks, posing a serious threat to intellectual property (IP). Third-party ownership verification provides an effective mechanism to protect IP of model owners while ensuring impartial verification. Current GNN fingerprint verification methods impose impractical requirements, demanding either access to internal model fingerprints (e.g., node embeddings) or detailed information about model stealing attacks. In this paper, we propose Canary, a black-box third-party GNN ownership verification protocol based on fingerprints, where verifiers access only model prediction posteriors and require no knowledge of adversaries' model stealing attacks. Specifically, we generate subgraph inputs as decision boundary fingerprints by maximizing the logits distance between the shadow surrogate and shadow independent model, effectively revealing their posterior probability discrepancies. Experiments across six datasets show that Canary effectively detects ten types of model stealing attacks, achieving a 35.68% higher negative-class F1-score than the state-of-the-art gray-box method Grove while maintaining robustness against seven types of evasion attacks.

Machine learning (ML) is promising in accurately detecting malicious flows in encrypted network traffic; however, it is challenging to collect a training dataset that contains a sufficient amount of encrypted malicious data with correct labels. When ML models are trained with low-quality training data, they suffer degraded performance. In this paper, we aim to address a real-world low-quality training dataset problem, namely, detecting encrypted malicious traffic generated by continuously evolving malware. We develop RAPIER+ that fully utilizes different distributions of normal and malicious traffic data in the feature space, where normal data is tightly distributed in a certain area, and the malicious data is scattered over the entire feature space to augment training data for model training. RAPIER+ includes two pre-processing modules to convert traffic into feature vectors and correct label noises. We evaluate our system on two public datasets and one combined dataset. With 1000 samples and 45% noise from each dataset, our system achieves the F1 scores of 0.78, 0.84, and 0.87, respectively, achieving average improvements of 358.5%, 314.0%, and 221.1% over the existing methods, respectively. Furthermore, we evaluate RAPIER+ with a real-world dataset obtained from a security enterprise. RAPIER+ effectively achieves encrypted malicious traffic detection with the best F1 score of 0.81 and improves the F1 score of existing methods by an average of 288.7%.