Attack Detection and Investigation
What happened, and why? A trustworthy record can still contain millions of ordinary events, and the question is which connections matter. We turn audit logs into provenance graphs, which record how processes, files, and network connections affect one another, and we detect and investigate attacks on those graphs. FLASH (IEEE S&P 2024) learns benign behavior with graph neural networks and flags activity that does not fit. HADES traces Active Directory attacks across machines, and CPD detects attackers who set up persistence and return later. For analysts, NoDoze ranks alerts by how rare their causal history is, and our small language models turn plain questions into KQL threat-hunting queries. RealityCheck carries the same investigation to AR/VR headsets, where standard auditing produces incomplete graphs or does not run at all.
Selected Papers
2026
- RAIDEffective and Efficient Threat Hunting with Small Language ModelsSaleha Muzammil, Rahul Reddy, Vishal Kamalakrishnan, Hadi Ahmadi, and Wajih Ul HassanIn International Symposium on Research in Attacks, Intrusions, and Defenses (RAID), 2026
Analysts in Security Operations Centers query massive telemetry streams using Kusto Query Language (KQL), but writing correct KQL demands specialized expertise that bottlenecks scaling security teams. We investigate how Small Language Models (SLMs) can enable accurate, cost-effective translation from natural language queries (NLQs) to KQL. We propose a three-knob framework spanning prompting, fine-tuning, and architecture. First, we adapt NL2KQL for SLMs with lightweight retrieval and introduce error-aware prompting that targets common parser failures with a handful of mined tips, at a fraction of the tokens KQL’s full rule set would require. Second, we apply LoRA fine-tuning with rationale distillation, augmenting each NLQ-KQL pair with a brief chain-of-thought to transfer teacher reasoning. This yields an informative negative result, as neither variant surpasses targeted prompting. Third, we propose a two-stage architecture pairing an SLM drafter with a low-cost LLM judge for schema-aware refinement. We evaluate nine models (five SLMs, four LLMs) on syntax correctness, semantic accuracy, table selection, filter precision, latency, and token cost. On Microsoft’s NL2KQL Defender Evaluation dataset, our two-stage approach reaches 0.987 syntax and 0.906 schema-valid ("semantic") accuracy, exceeding every baseline we run under equivalent infrastructure, and it generalizes to independently authored queries over the same schema (0.964 syntax, 0.831 schema-valid). The only baselines within 0.05 schema-valid are NL2KQL+GPT-4o (0.878) and NL2KQL+GPT-5 (0.861), which cost 2.998 and 2.018 for 230 queries against $0.213 for ours, a 9.5-14× reduction at matched accuracy. These results establish SLMs as a practical foundation for natural-language querying in security operations.
@inproceedings{muzammil2025towards, title = {Effective and Efficient Threat Hunting with Small Language Models}, author = {Muzammil, Saleha and Reddy, Rahul and Kamalakrishnan, Vishal and Ahmadi, Hadi and Hassan, Wajih Ul}, booktitle = {International Symposium on Research in Attacks, Intrusions, and Defenses (RAID)}, year = {2026}, } - TDSCAccurate and Scalable Detection and Investigation of Cyber Persistence ThreatsQi Liu, Muhammad Shoaib, Mati Ur Rehman, Kaibin Bao, Veit Hagenmeyer, and Wajih Ul HassanTransactions on Dependable and Secure Computing, 2026
In Advanced Persistent Threat (APT) attacks, achieving stealthy persistence within target systems is often crucial for an attacker’s success. This persistence allows adversaries to maintain prolonged access, often evading detection mechanisms. Recognizing its pivotal role in the APT lifecycle, this paper introduces Cyber Persistence Detector (CPD), a novel system dedicated to detecting cyber persistence through provenance analytics. CPD is founded on the insight that persistent operations typically manifest in two phases: the “persistence setup” and the subsequent “persistence execution”. By causally relating these phases, we enhance our ability to detect persistent threats. First, CPD discerns setups signaling an impending persistent threat and then traces processes linked to remote connections to identify persistence execution activities. A key feature of our system is the introduction of pseudo-dependency edges (pseudoedges), which effectively connect these disjoint phases using data provenance analysis, and expert-guided edges, which enable faster tracing and reduced log size. These edges empower us to detect persistence threats accurately and efficiently. Moreover, we propose a novel alert triage algorithm that further reduces false positives associated with persistence threats. Evaluations conducted on well-known datasets demonstrate that our system reduces the average false positive rate by 93% compared to state-of-the-art methods.
@article{liutdsc2026, title = {Accurate and Scalable Detection and Investigation of Cyber Persistence Threats}, author = {Liu, Qi and Shoaib, Muhammad and Rehman, Mati Ur and Bao, Kaibin and Hagenmeyer, Veit and Hassan, Wajih Ul}, journal = {Transactions on Dependable and Secure Computing}, year = {2026}, doi = {10.1109/tdsc.2026.3689905} }
2025
- USENIX SecurityPrincipled and Automated Approach for Investigating AR/VR AttacksMuhammad Shoaib, Alex Suh, and Wajih Ul HassanIn USENIX Security Symposium, 2025
As Augmented and Virtual Reality (AR/VR) adoption grows across sectors, auditing systems are needed to enable provenance analysis of AR/VR attacks. However, traditional auditing systems often generate inaccurate and incomplete provenance graphs, or fail to work due to operational restrictions in AR/VR devices. This paper presents REALITYCHECK, a provenance-based auditing system designed to support accurate root cause analysis and impact assessments of complex AR/VR attacks. Our system first enhances the W3C PROV data model with additional ontology to capture AR/VR-specific entities and causal relationships. Then, we employ a novel adaptation of natural language processing and feature-based log correlation techniques to transparently extract entities and relationships from dispersed, unstructured AR/VR logs into provenance graphs. Finally, we introduce an AR/VR-aware execution partitioning technique to filter out forensically irrelevant data and false causal relationships from these provenance graphs, improving analysis accuracy and investigation speed. We built a REALITYCHECK prototype for Meta Quest 2 and evaluated it against 25 real-world AR/VR attacks. The results show that REALITYCHECK generates accurate provenance graphs for all AR/VR attacks and incurs low runtime overhead across benchmarked applications. Notably, our execution partitioning approach drastically reduces the size of the graph without sacrificing essential investigation details. Our system operates non-intrusively, requires no additional installation, and is generalizable across various AR/VR devices.
@inproceedings{realitycheck, title = {Principled and Automated Approach for Investigating AR/VR Attacks}, author = {Shoaib, Muhammad and Suh, Alex and Hassan, Wajih Ul}, year = {2025}, booktitle = {USENIX Security Symposium}, } - TDSCHADES: Detecting Active Directory Attacks via Whole Network Provenance AnalyticsQi Liu, Kaibin Bao, Wajih Ul Hassan, and Veit HagenmeyerIEEE Transactions on Dependable and Secure Computing, 2025
Due to its crucial role in identity and access management in modern enterprise networks, Active Directory (AD) is a top target of Advanced Persistence Threat (APT) actors. Conventional intrusion detection systems (IDS) excel at identifying malicious behaviors caused by malware, but often fail to detect stealthy attacks launched by APT actors. Recent advance in provenance-based IDS (PIDS) shows promises by exposing malicious system activities in causal attack graphs. However, existing approaches are restricted to intra-machine tracing, and unable to reveal the scope of attackers’ traversal inside a network. We propose HADES, the first PIDS capable of performing accurate causality-based cross-machine tracing by leveraging a novel concept called logon session based execution partitioning to overcome several challenges in cross-machine tracing. We design HADES as an efficient on-demand tracing system, which performs whole-network tracing only when it first identifies an authentication anomaly signifying an ongoing AD attack, for which we introduce a novel lightweight authentication anomaly detection model rooted in our extensive analysis of AD attacks. To triage attack alerts, we present a new algorithm integrating two key insights we identified in AD attacks. Our evaluations show that HADES outperforms both popular open-source detection systems and a prominent commercial AD attack detector.
@article{liu2025hades, title = {HADES: Detecting Active Directory Attacks via Whole Network Provenance Analytics}, author = {Liu, Qi and Bao, Kaibin and Hassan, Wajih Ul and Hagenmeyer, Veit}, journal = {IEEE Transactions on Dependable and Secure Computing}, year = {2025}, doi = {10.1109/tdsc.2025.3611866} }
2024
- IEEE S&PFLASH: A Comprehensive Approach to Intrusion Detection via Provenance Graph Representation LearningMati Ur Rehman, Hadi Ahmadi, and Wajih Ul HassanIn IEEE Symposium on Security and Privacy (S&P), 2024
Recently, provenance-based Intrusion Detection Systems (IDSes) have gained popularity for their potential in detecting sophisticated Advanced Persistent Threat (APT) attacks. These IDSes employ provenance graphs created from system logs to identify potentially malicious activities. Despite their potential, they face challenges in accuracy, practicality, and scalability, particularly when dealing with large provenance graphs. We present Flash, a scalable IDS that leverages graph representation learning through Graph Neural Networks (GNNs) on data provenance graphs to overcome these limitations. Flash employs a Word2Vec-based semantic encoder to capture essential semantic attributes (e.g., process names and file paths) and the temporal ordering of events within the provenance graph. Furthermore, Flash incorporates a novel adaptation of a GNN-based contextual encoder to efficiently encode both local and global graph structures into expressive node embeddings. To learn benign node behaviors, we utilize a lightweight classifier that combines the GNN and Word2Vec embeddings. Recognizing the computational demands and slow processing times of GNN, particularly for large provenance graphs, we have developed an embedding recycling database to store the node embeddings generated during the training phase. During runtime, our lightweight classifier leverages the stored embeddings, obviating the need to regenerate GNN embeddings, thus facilitating real-time APT detection. Extensive evaluation of Flash on real-world datasets demonstrates superior detection accuracy compared to existing provenance-based IDSes. The results also illustrate Flash’s scalability, robustness against mimicry attacks, and potential for accelerating the alert verification process.
@inproceedings{flash2024, title = {FLASH: A Comprehensive Approach to Intrusion Detection via Provenance Graph Representation Learning}, author = {Rehman, Mati Ur and Ahmadi, Hadi and Hassan, Wajih Ul}, booktitle = {IEEE Symposium on Security and Privacy (S\&P)}, year = {2024}, doi = {10.1109/sp54263.2024.00139} }
2019
- NDSSNoDoze: Combatting threat alert fatigue with automated provenance triageWajih Ul Hassan, Shengjian Guo, Ding Li, Zhengzhang Chen, Kangkook Jee, Zhichun Li, and Adam BatesIn Symposium on Network and Distributed System Security (NDSS), 2019
Large enterprises are increasingly relying on threat detection softwares (e.g., Intrusion Detection Systems) to allow them to spot suspicious activities. These softwares generate alerts which must be investigated by cyber analysts to figure out if they are true attacks. Unfortunately, in practice, there are more alerts than cyber analysts can properly investigate. This leads to a "threat alert fatigue" or information overload problem where cyber analysts miss true attack alerts in the noise of false alarms. In this paper, we present NODOZE to combat this challenge using contextual and historical information of generated threat alert. NODOZE first generates a causal dependency graph of an alert event. Then, it assigns an anomaly score to each edge in the dependency graph based on the frequency with which related events have happened before in the enterprise. NODOZE then propagates those scores along the neighboring edges of the graph using a novel network diffusion algorithm and generates an aggregate anomaly score which is used for triaging. We deployed and evaluated NODOZE at NEC Labs America. Evaluation on our dataset of 364 threat alerts shows that NODOZE consistently ranked the true alerts higher than the false alerts based on aggregate anomaly scores. Further, through the introduction of a cutoff threshold for anomaly scores, we estimate that our system decreases the volume of false alarms by 84%, saving analysts’ more than 90 hours of investigation time per week. NODOZE generates alert dependency graphs that are two orders of magnitude smaller than those generated by traditional tools without sacrificing the vital information needed for the investigation. Our system has a low average runtime overhead and can be deployed with any threat detection software.
@inproceedings{nodoze2019, title = {{NoDoze}: Combatting threat alert fatigue with automated provenance triage}, author = {Hassan, Wajih Ul and Guo, Shengjian and Li, Ding and Chen, Zhengzhang and Jee, Kangkook and Li, Zhichun and Bates, Adam}, year = {2019}, booktitle = {Symposium on Network and Distributed System Security (NDSS)}, doi = {10.14722/ndss.2019.23349} }