The 30th ACM Symposium on Access Control Models and Technologies

Bottom-up role-mining is the determination of a set of roles given as input a set of users and the permissions those users possess. It is well-established in the research literature, and in practice, as an important problem in information security. A natural objective that has been explored in prior work is for the set of roles to be of minimum size. We address this problem for practical inputs while reconciling foundations, specifically, that the problem is NP-hard. We first observe that an approach from prior work that exploits a sufficient condition for an efficient algorithm, while a useful first step, does not scale to more recently proposed benchmark inputs. We propose a new technique: the enumeration of maximal bicliques. We point out that the number of maximal bicliques provides a natural measure of the hardness of an input. We leverage the enumeration of maximal bicliques in two different ways. Our first approach addresses more than half the benchmark inputs to yield exact results. The other approach is the basis for a heuristic for hard instances; in it, we identify and adopt as roles those that correspond to large maximal bicliques. We have implemented all our algorithms and carried out an extensive empirical assessment, which suggests that our approaches are promising. We make our code available publicly as open-source.

Access control mechanisms are essential for securing organizational resources, with Attribute-based access control (ABAC) offering fine-grained and dynamic policy enforcement. However, traditional ABAC implementations often rely on static rule sets, which may become outdated as organizational structures and user behaviors evolve. This article introduces a predictive model that enhances ABAC policies by leveraging historical access logs and machine-learning techniques. By analyzing past access patterns, our approach refines access control decisions, dynamically adjusting policies based on user behavior and contextual insights. The proposed model employs hierarchical clustering to detect behavioral similarities and anomalies, improving policy adaptation and reducing misclassifications in access requests. Experimental validation using synthetic datasets demonstrates that this predictive enhancement significantly improves access control accuracy, minimizes unauthorized access risks, and reduces administrative overhead. Our findings highlight the potential of data-driven approaches in optimizing ABAC policies for evolving security requirements.

Role mining is the problem of translating a user-permission access matrix into a Role-Based Access Control (RBAC) policy. We address role mining for the minimum number of edges in the output RBAC policy, which is a natural, quantitative objective that has been explored in prior work. The problem is known to be intractable (NP-hard); we seek approaches that work well for instances that arise in practice. We first revisit approaches from prior work, for all of which we provide implementations. We find that for benchmark inputs, only one approach scales, which we adopt as our baseline. We then propose two new approaches. One is an improvement over a previously proposed approach of greedily choosing a largest possible biclique in the input access matrix. The other leverages a technique for minimizing the number of roles, and reduces the number of edges in that solution by considering subsets of roles at a time. We have carried out a thorough empirical assessment against two sets of benchmarks for role mining. We make all our code available as open-source.

Hardware capabilities offer an alternative to how access control is commonly implemented in processors today, i.e., through enforcement of permission checks on virtual memory at the time of address translation. Despite conceptual strengths of capability hardware, it is challenging for existing capability-oblivious software stacks to be compatible with capability hardware and to interoperate with capability-aware software, hindering faster adoption. Prior attempts to achieve this sacrifice the inherent advantages of capabilities. They require trusting a software central authority (e.g., the OS kernel) for capability-based isolation and limit the scope of capability-based memory sharing to individual virtual address spaces. This paper proposes the idea of caplification, a novel mechanism to enable seamless co-existence of capability-aware and capability-oblivious software stacks. We concretely implement our proposed idea on a modern RISC-V capability hardware and show how it enables running a commodity unmodified (or capability-oblivious) Linux OS. Our design retains the full advantages provided by hardware capabilities, such as creating fine-grained hardware-isolated memory compartments both in user- and kernel-space. We evaluated our prototype system both on QEMU emulation and on hardware RTL simulation. We find that the performance of our system is comparable to prior baseline designs, while offering cost improvements in scenarios of secure data sharing.

Cryptographic Access Control (CAC) allows organizations to control cloud-hosted data sharing among users while preventing external attackers, malicious insiders, and honest-but-curious cloud providers from accessing the data. However, CAC entails an overhead often impractical for real-world scenarios due to the many cryptographic computations involved. Hence, we put forth a hybrid Access Control (AC) scheme --- combining CAC and (traditional) centralized AC --- that considers trust assumptions (e.g., on users) and data protection requirements of the underlying scenario on a case-by-case basis to reduce the number of cryptographic computations to execute in CAC. Besides, we design a consistency check to ensure the correctness and safety properties of the enforcement of the hybrid AC scheme, provide a proof-of-concept implementation in Prolog, and conduct a preliminary experimental evaluation.

Access and usage control have evolved to include obligations, which are mandatory actions that must be fulfilled as part of authorization decisions. However, standards such as Abbreviated Language For Authorization (ALFA) and eXtensible Access Control Markup Language (XACML) specify that Policy Enforcement Points (PEPs) are responsible for enforcing obligations but leave execution aspects unspecified. They assume that obligations will be fulfilled without addressing how enforcement should be carried out. This paper introduces an enforcement framework based on structured enforcement messages to address these challenges. It defines two types of messages to coordinate enforcement execution across PEPs: Declaration and Execution Records. These records provide information about PEP capabilities, action dependencies, and fallback strategies for synchronized enforcement. Secondly, we propose a hierarchical policy model to separate concerns between the functionality of policy logic and enforcement. The model is composed of Governance, Authorization, and Enforcement Policies. Constraints flow across these three levels, allowing policy selection and execution to adapt to the authorization context and enforcement. Finally, we illustrate a high-level architecture that integrates the three policy layers with distributed enforcement logic across multiple PEPs.

Graph-based frameworks are often used in network hardening to help a cyber defender understand how a network can be attacked and how the best defenses can be deployed. However, incorporating network connectivity parameters in the attack graph, reasoning about the attack graph when we do not have access to complete information, providing system administrator suggestions in an understandable format, and allowing them to do what-if analysis on various scenarios and attacker motives is still missing. We fill this gap by presenting SPEAR, a formal framework with tool support for security posture evaluation and analysis that keeps human-in-the-loop. SPEAR uses the causal formalism of AI planning to model vulnerabilities and configurations in a networked system. It automatically converts network configurations and vulnerability descriptions into planning models expressed in the Planning Domain Definition Language (PDDL). SPEAR identifies a set of diverse security hardening strategies that can be presented in a manner understandable to the domain expert. These allow the administrator to explore the network hardening solution space in a systematic fashion and help evaluate the impact and compare the different solutions.

Recent revelations of various access control vulnerabilities in 5G core networks have raised severe concerns about their security. Unfortunately, these vulnerabilities are difficult to patch, as they stem from ambiguity and under-specification in 3GPP specifications. This research introduces a new practical solution called ACGuard5GC (Access Control Guard for 5G Core Networks) to prevent the potential attacks exploiting these access control vulnerabilities in 5G core networks. Its key idea is to deploy reference monitors alongside the Service Communication Proxies (SCPs), which facilitate communications across different network functions in 5G core networks as well as enhancing their scalability and observability. The reference monitors in ACGuard5GC enforce multiple safety properties to prevent the existing access control attacks discovered in the literature. To address the privacy concern in distributed and cooperative defenses by multiple SCPs, ACGuard5GC applies private set intersection on information exchanged among different SCPs. For performance evaluation, we provide a reference implementation of SCP and instrument it with reference monitor functionalities that enforce different safety properties. Our experimental results with two existing 5G core network emulators, Open5GS and VET5G, demonstrate that ACGuard5GC is capable of preventing access control attacks effectively while incurring low operational overhead.

Cellular networks require strict security procedures and measures across various network components, from core to radio access network (RAN) and end-user devices. As networks become increasingly complex and interconnected, as in O-RAN deployments, they are exposed to a numerous security threats. Therefore, ensuring robust security is critical for O-RAN to protect network integrity and safeguard user data. This requires rigorous testing methodologies to mitigate threats. This paper introduces an automated, adaptive, and scalable user equipment (UE) based RAN security testing framework designed to address the shortcomings of existing RAN testing solutions. Experimental results on a 5G software radio testbed built with commercial off-the-shelf hardware and open source software validate the efficiency and reproducibility of sample security test procedures developed on the RAN Tester UE framework.

🥈 Best Paper Runner-Up Award

We study the safety problem for the next-generation access control (NGAC) model. We show that under mild assumptions it is coNP-complete, and under further realistic assumptions we give an algorithm for the safety problem that significantly outperforms naive brute force search. We also show that real-world examples of mutually exclusive attributes lead to nearly worst-case behavior of our algorithm.

Enterprises that own or handle sensitive resources rely on access control models to protect those resources. NIST has recently standardized a new access control model called Next Generation Access Control (NGAC) and provided a reference implementation. Despite the importance of properly functioning access control systems, little work has been done to verify that the software implementing NGAC properly conforms to the NGAC standard. Prior approaches for finding bugs are either designed to detect fail-stop faults, or model the protocol or software itself, which does not identify discrepancies between software and standards.

In this paper, we solve this problem with a methodology we call policy engine differential mutation analysis, which we implement in a system called ProfessorX. ProfessorX detects access decision discrepancies between policy engine implementations—specifically, the NIST reference implementation, and our own implementation of the standard. If there are no discrepancies, we mutate the policy slightly and then try again. Using this technique, we identified two novel vulnerabilities, and show that our system is fast enough to be practically useful to developers.

Attribute-Based Access Control (ABAC) provides expressiveness and flexibility, making it a compelling model for enforcing fine-grained access control policies. To facilitate the transition to ABAC, extensive research has been conducted to develop methodologies, frameworks, and tools that assist policy administrators in adapting the model. Despite these efforts, challenges remain in the availability and benchmarking of ABAC datasets. Specifically, there is a lack of clarity on how datasets can be systematically acquired, no standardized benchmarking practices to evaluate existing methodologies and their effectiveness, and limited access to real-world datasets suitable for policy analysis and testing.

This paper introduces ABAC Lab, an interactive platform that addresses these challenges by integrating existing ABAC policy datasets with analytical tools for policy evaluation. Additionally, we present two new ABAC datasets derived from real-world case studies. ABAC Lab serves as a valuable resource for both researchers studying ABAC policies and policy administrators seeking to adopt ABAC within their organizations. By offering an environment for dataset exploration and policy analysis, ABAC Lab facilitates research, aids policy administrators in transitioning to ABAC, and promotes a more structured approach to ABAC policy evaluation and development.

🥇 Best Paper Award

Access control (AC) systems play an important role in ensuring security by regulating how resources are accessed, protecting sensitive information, and maintaining system integrity. Their complexity arises not only from diverse policies and mechanisms but also from the involvement of multiple stakeholders, including resource owners, administrators, and end-users. Taking inspiration from explainable AI and explainable security, we define the first model of access control explainability, as a quality measure of the explanation graph constructed around the decisions made within the AC system. We then explore the literature to identify how existing work can be integrated as explanatory processes. Finally, we leverage our framework to articulate three open research challenges: the collection and interpretation of AC decisions, the effective construction of AC explanation graphs, and the definition of meaningful and computationally efficient explanation quality metrics.

In the Ethereum system, the exclusion of specific transactions is currently feasible with minimal effort due to a power imbalance among entities. This censorship opportunity threatens the dependability of time-sensitive services deployable on Ethereum. In this paper, we look at this threat from an access control perspective and attribute it to a lack of accountability for censorship, a lack of policy definition and enforcement, as well as to the lack of disincentivization of policy violation. We propose an approach to enforceable policies in Ethereum. Furthermore, we demonstrate how a specific policy can address the shortcomings of existing censorship mitigation techniques, particularly Inclusion Lists. Under the assumption that block assemblers are unwilling to incur significant financial penalties as well as that the local view on the network messages is sufficiently consistent, the proposed approach guarantees the inclusion of a transaction in a block within 27 seconds in a non-saturated network. The empirical validation of sufficiently consistent views on outstanding transactions is currently in progress.

Easily Deployable and Efficiently Searchable Encryption (EDESE) is a cryptographic primitive designed for practical searchable applications, offering efficient search and easy deployment. However, it remains vulnerable to Leakage-Abuse attacks, allowing adversaries to exploit keyword-matching processes to extract sensitive information. To address these vulnerabilities, we introduce Leakage-Resilient EDESE (LR-EDESE) with k-indistinguishability and controlled leakage functions. We then propose Volume Leakage-Resilient EDESE (VLR-EDESE), a new scheme to protect against both query and document volume leakage. Our experimental results demonstrate that at k = 5000 (maximum security setting), VLR-EDESE incurs an overhead of 63× compared to the baseline EDESE without leakage protection, outperforming state-of-the-art methods with 320× and 97× overhead, respectively. For smaller k values (10, 20, 50, 100), storage and communication overhead remain within 2× and 2.5× of the baseline EDESE, highlighting VLR-EDESE's flexibility. Finally, we present CloudSec, an implementation of VLR-EDESE that seamlessly integrates with cloud storage platforms, using OneDrive as an example.

Developing and managing access control systems is challenging due to the dynamic nature of users, resources, and environments. Recent advancements in machine learning (ML) offer promising solutions for automating the extraction of access control attributes, policy mining, verification, and decision-making. Despite these advancements, the application of ML in access control remains fragmented, resulting in an incomplete understanding of best practices. This work aims to systematize the use of ML in access control by identifying key components where ML can address various access control challenges. We propose a novel taxonomy of ML applications within this domain, highlighting current limitations such as the scarcity of public real-world datasets, the complexities of administering ML-based systems, and the opacity of ML model decisions. Additionally, we outline potential future research directions to guide both new and experienced researchers in effectively integrating ML into access control practices.

Natural Language Access Control Policies (NLACPs) define who can access specific information within an organization and under what conditions. While these policies are typically written in semi-formal or informal natural language, making them easily interpretable by humans, they cannot be directly enforced by access control systems. Their unstructured nature introduces ambiguities and inconsistencies, making automated extraction and translation into structured, machine-enforceable security rules a significant challenge.

This paper presents a two-step LLM-based approach for transforming security policies expressed in natural language text (NLT) into machine-enforceable ABAC rules. Our approach first classifies policy statements to determine whether they constitute an NLACP. If a statement is policy-related, it is then translated into a structured ABAC (Attribute-Based Access Control) rule, extracting conditions of subject attributes, object attributes, and actions. To enhance accuracy and efficiency while reducing reliance on human annotations, we employ a two-tiered knowledge distillation strategy, where GPT-4 generates synthetic training data to fine-tune a lightweight CodeT5 student model. We evaluate our method on real-world datasets, demonstrating high classification accuracy and effective structured policy extraction. Our findings indicate that LLM-driven structured policy conversion enables scalable, automated integration of NLACPs directly into an access control system, thereby reducing manual effort in translating a natural language text into an ABAC security policy.

The widespread adoption of Large Language Models (LLMs) in software development has accelerated coding workflows, with tools like GitHub Copilot and Google Gemini reducing development time by up to 55.8%. However, these systems suffer from security and trustworthiness challenges. In this paper, we demonstrate a novel attack - ''Vulnerability propagation attack'' in the conext of the code generated by LLMs. We present a formal framework for evaluating vulnerability propagation in LLM-assisted development, focusing on two key metrics: Vulnerability Carry Rate (VCR), which quantifies the intra-session propagation of vulnerabilities, and Context Retention Factor (CRF), which measures cross-session persistence. Using three leading LLMs -- GPT-4, Sonar, and Gemini -- we evaluate 10 critical Common Weakness Enumeration (CWE) categories across three tiers of prompts (direct injection, implicit vulnerabilities, and adversarial few-shot prompting). Our findings reveal that adversarial few-shot prompting exacerbates vulnerability propagation, with VCR values reaching 100% under high-difficulty conditions and CRF values demonstrating significant retention across sessions. These results underscore the need for session-aware security measures and robust vulnerability detection frameworks in AI-assisted development environments.

Recent legal frameworks have mandated the right to be forgotten, obligating the removal of specific data upon user requests. Machine Unlearning has emerged as a promising solution by selectively removing learned information from machine learning models. This paper presents MUBox, a comprehensive platform designed to evaluate unlearning methods in deep learning. MUBox integrates 23 advanced unlearning techniques, tested across six practical scenarios with 11 diverse evaluation metrics. It allows researchers and practitioners to (1) assess and compare the effectiveness of different machine unlearning methods across various scenarios; (2) examine the impact of current evaluation metrics on unlearning performance; and (3) conduct detailed comparative studies on machine unlearning in a unified framework. Leveraging MUBox, we systematically evaluate these unlearning methods in deep learning and uncover a set of key insights: (a) Even state-of-the-art unlearning methods, including those published in top-tier venues and winners of unlearning competitions, demonstrate inconsistent effectiveness across diverse scenarios. Prior research has predominantly focused on simplified settings, such as random forgetting and class-wise unlearning, highlighting the need for broader evaluations across more complex and realistic unlearning tasks. (b) Assessing unlearning performance remains a non-trivial problem, as no single evaluation metric can comprehensively capture the effectiveness, efficiency, and preservation of model utility. Our findings emphasize the necessity of employing multiple metrics to achieve a balanced and holistic assessment of unlearning methods. (c) In the context of depoisoning-removing the adverse effects of poisoned data-our evaluation reveals significant variability in the effectiveness of existing approaches, which is highly dependent on the specific type of poisoning attack. We believe MUBox will serve as a valuable benchmark, advancing research in machine unlearning and highlighting areas for future improvement. Codes are available at https://github.com/Jessegator/MUBox.

Geofencing, the virtual demarcation of physical spaces, is widely used for managing the localisation of Internet of Things (IoT) devices. However, traditional localisation techniques face security challenges indoors due to signal interference and susceptibility to spoofing, often requiring extensive calibration or extra hardware, limiting scalability. In this work, we propose ZeroTouch, a machine learning-based system that leverages Received Signal Strength (RSS) measurements from multiple receivers to improve the security of geofencing without introducing additional deployment overhead. While RSS-based localisation is known to have inherent security limitations, we show that by aggregating RSS readings from multiple anchor points and detecting anomalies using an autoencoder model, ZeroTouch provides a practical and automated mechanism for verifying whether a device is inside or outside a defined boundary. Rather than serving as a standalone security mechanism, ZeroTouch enhances existing authentication frameworks by adding an additional zero-touch security layer that operates passively in the background. ZeroTouch eliminates manual calibration, removes the human-in-the-loop element, and simplifies deployment. We evaluate our solution in a realistic simulated environment and demonstrate that it achieves high accuracy in distinguishing between in-room and out-of-room devices, even in strong adversarial settings.

Internet-of-Things (IoT) is playing a key role in modern society by offering enhanced functionalities and services. As IoT devices may introduce new security risks to the network, network administrators profile the behavior of IoT devices using device fingerprinting. Device fingerprinting typically involves training a machine learning model using the network behavioral data of existing devices. If a new device is added, the network becomes vulnerable to attacks until the time that the machine learning model is trained and updated to integrate the new device. Furthermore, if many devices are regularly added to the network, the cost of adapting the machine learning model can be significant. To address the challenges of security and scalability in fingerprinting, we create a collection of observed behaviors of IoT devices from existing devices and use this collection to construct a fingerprint for a new device. In our approach, we design a bi-component neural network architecture consisting of a transformer-based behavior-extractor (BE) and a fingerprinting interpreter.We perform a one-time training of the BE to extract behaviors from known devices. We use the generated BE for (a) fingerprinting existing devices and (b) adapting the existing fingerprinting model to new device data. In our experiments on 22 diverse IoT devices, we show that our model can identify newly introduced devices as well as known devices with a high identification rate. Our approach improves the time to adapt a model by a factor of 78.3× with no loss of accuracy, achieving recall over 98%.