arXiv subjects

Chenang Li

Publications and source records attributed to Chenang Li.

3 recordsLinked to original sources

arXiv2026

Selection-Aware Stress Testing for Interactive Agents

Agent evaluations often use one benchmark to choose a workflow and then search for task types where its advantage weakens, so both conclusions are selected from the same data. We introduce Selection-Aware Semantic Stress Testing (\SASST{}), which learns a task reweighting from pre-execution features on discovery tasks and evaluates the same paired comparison on separate confirmation tasks. The protocol checks support and stability, uses joint bounds for all planned claims, and can return no claim. We prove conditional asymptotic validity under stated cluster assumptions. A forty-cluster audit finds Gaussian undercoverage and conservative Bonferroni $t$ bounds. In one 480-episode $\tau$-bench study, a $3.75$ point discovery gain vanished on confirmation. A second-model study likewise confirmed neither a workflow benefit nor a stable stress rule.

cs.LG↗

arXiv2025

Convergent Privacy Framework for Multi-layer GNNs through Contractive Message Passing

Differential privacy (DP) has been integrated into graph neural networks (GNNs) to protect sensitive structural information, e.g., edges, nodes, and associated features across various applications. A prominent approach is to perturb the message-passing process, which forms the core of most GNN architectures. However, existing methods typically incur a privacy cost that grows linearly with the number of layers (e.g., GAP published in Usenix Security'23), ultimately requiring excessive noise to maintain a reasonable privacy level. This limitation becomes particularly problematic when multi-layer GNNs, which have shown better performance than one-layer GNN, are used to process graph data with sensitive information. In this paper, we theoretically establish that the privacy budget converges with respect to the number of layers by applying privacy amplification techniques to the message-passing process, exploiting the contractive properties inherent to standard GNN operations. Motivated by this analysis, we propose a simple yet effective Contractive Graph Layer (CGL) that ensures the contractiveness required for theoretical guarantees while preserving model utility. Our framework, CARIBOU, supports both training and inference, equipped with a contractive aggregation module, a privacy allocation module, and a privacy auditing module. Experimental evaluations demonstrate that CARIBOU significantly improves the privacy-utility trade-off and achieves superior performance in privacy auditing tasks.

cs.CR↗

arXiv2025

Entente: Cross-silo Intrusion Detection on Network Log Graphs with Federated Learning

Graph-based Network Intrusion Detection Systems (GNIDS) have gained significant momentum in detecting sophisticated cyber-attacks, such as Advanced Persistent Threats (APTs), within and across organizational boundaries. Though achieving satisfying detection accuracy and demonstrating adaptability to ever-changing attacks and normal patterns, existing GNIDS predominantly assume a centralized data setting. However, flexible data collection is not always realistic or achievable due to increasing constraints from privacy regulations and operational limitations. We argue that the practical development of GNIDS requires accounting for distributed collection settings and we leverage Federated Learning (FL) as a viable paradigm to address this prominent challenge. We observe that naively applying FL to GNIDS is unlikely to be effective, due to issues like graph heterogeneity over clients and the diverse design choices taken by different GNIDS. We address these issues with a set of novel techniques tailored to the graph datasets, including reference graph synthesis, graph sketching and adaptive contribution scaling, eventually developing a new system Entente. By leveraging the domain knowledge, Entente can achieve effectiveness, scalability and robustness simultaneously. Empirical evaluation on the large-scale LANL, OpTC and Pivoting datasets shows that Entente outperforms the SOTA FL baselines. We also evaluate Entente under FL poisoning attacks tailored to the GNIDS setting, showing the robustness by bounding the attack success rate to low values. Overall, our study suggests a promising direction to build cross-silo GNIDS.

cs.CR↗