Software supply chain security is an essential area of cybersecurity, as shown by attacks such as the XZ Utils incident in 2024. Software Bills of Materials (SBOMs) were introduced to keep track of the supply chain of a software artifact. They have been brought to widespread attention in 2021 through US Executive Order 14028. The EU will make SBOMs mandatory by 2027 with their Cyber Resilience Act from 2024. In contrast to these demands for SBOMs through legislation, practitioners struggle to make real use of them. Multiple recent studies have concluded that adoption of SBOMs and their integration into security processes are facing various challenges. At the same time, accuracy and correctness problems of SBOMs generated by wide-spread tools have been shown. Based on this situation, the dissertation research is dedicated to the research question: What benefit do SBOMs provide and what further benefit can they potentially provide for supply chain security? In order to answer, this research description introduces four intended contributions: (1) First, all knowledge about the history of SBOMs and related concepts in Computer Science is systematized. (2) Second, an SBOM usage model that provides an abstract view on the software supply chain as well as actors involved in it is developed. The goal of the model is to give a theoretical foundation for making practical use of SBOMs. (3) Third, with BOM2VULN an analysis of current tools that map vulnerabilities to SBOMs is conducted, potentially including the introduction of a new tool for this task. The goal is to provide security engineers with a means of quickly finding vulnerabilities and assessing their relevance. (4) Lastly, the SBOM Nutri Score is proposed including an evaluation. The score helps practitioners to evaluate the software supply chain risk of third party code they use or intend to use.
Man-At-The-End (MATE) attackers operate with full access to software or hardware targets and can observe, analyze, and modify running systems to extract secrets or alter behavior. CheckMATE explores both offensive and defensive research in this space: measurement studies and tooling that expose realistic attack techniques, alongside defenses such as obfuscation, tamper-resistance, watermarking, white-box cryptography, and hardware-assisted protections. This workshop collects rigorous, reproducible research aimed at bridging academic advances and industry practice. The CheckMATE '25 complete workshop proceedings can be found at: https://dl.acm.org/citation.cfm?id=3733817
Most homomorphic encryption (FHE) schemes exploit a technique called single-instruction multiple-data (SIMD) to process several messages in parallel. However, they base their security in somehow strong assumptions, such as the hardness of approximate lattice problems with superpolynomial approximation factor. On the other extreme of the spectrum, there are lightweight FHE schemes that have much faster bootstrapping but no SIMD capabilities. On the positive side, the security of these schemes is based on lattice problems with (low-degree) polynomial approximation factor only, which is a much weaker security assumption. Aiming the best of those two options, Micciancio and Sorrell (ICALP'18) proposed a new amortized bootstrapping that can process many messages at once, yielding sublinear time complexity per message, and allowing one to construct FHE based on lattice problems with polynomial approximation factor. Some subsequent works on this line achieve near-optimal asymptotic performance, nevertheless, concrete efficiency remains mostly an open problem. The only existing implementation to date (GPV23, Asiacrypt 2023) requires keys of up to a hundred gigabytes while only providing gains for relatively large messages. In this paper, we introduce a new method for amortized bootstrapping where the number of homomorphic operations required per message is O(h) and the noise overhead is O(root h lambda log lambda), where h is the Hamming weight of the LWE secret key and lambda is the security parameter. This allows us to use much smaller parameters and to obtain faster running time. Our method is based on a new efficient homomorphic evaluation of sparse polynomial multiplication. We bootstrap 2 to 8-bit messages in 1.46 ms to 28.5 ms, respectively. Compared to TFHE-rs, this represents a performance improvement of 2.5 to 38.7 times while requiring bootstrapping keys up to 47.5 times smaller.
The widespread adoption of Large Language Models (LLMs) is driving the rapidly growing demand for large-scale computations like training and fine-tuning models. In many areas, the confidentiality of the underlying data is of critical importance to their corporate or government owners. However, securing data in large-scale computations is challenging. First, its demand for enormous hardware resources typically requires outsourcing (e.g., to the public cloud). Second, the large and rapidly evolving software stack used in LLM training in conjunction with a growing incidence of supply chain attacks and software vulnerabilities makes it all but impossible for data owners to establish trust in the code that processes their highly sensitive data. Confidential computing and sandboxing are promising techniques for solving these problems. However, existing sandboxes do not address covert channels which limits their ability to protect confidential data. This paper proposes IOVALVE, a novel I/O sandbox for large-scale computations on confidential data. IOVALVE places sandbox enforcement on a programmable network device that is physically isolated from the processor hardware running the untrusted software stack. This construction allows IOVALVE to sidestep the multitude of side channels due to visible or hidden resource sharing. IOVALVE interposes on all network I/O of the sandbox and only transmits encrypted and regularized network traffic in order to prevent information leakage over the network. Our evaluation shows that IOVALVE has marginal performance overhead and supports real-world applications like LLM fine-tuning and batch inference, and molecular simulation.
As the de-facto standard for container orchestration, Kubernetes is extensively adopted by numerous companies and cloud vendors, making its security critical. In this paper, we define a new attack surface called implicit permission: The execution of explicitly granted permissions in Kubernetes dynamically leads to implicit operations on other resources, enabling new permissions beyond the explicitly granted ones. Such implicit permissions create security vulnerabilities that attackers can exploit to compromise an entire cluster. Automatically identifying implicit permissions is challenging due to implicit relation reasoning and dynamic behaviors across diverse components of Kubernetes. To address that, we devise a systematic approach that combines static analysis techniques with the advanced capabilities of the large language model (LLM, e.g., GPT-4.5). Initially, we develop a static analysis to identify all Kubernetes resources. Building on this, we use static analysis to identify all explicit permissions for each resource. Finally, by combining the semantic reasoning capabilities of LLMs with the pattern-based precision of static analysis, we reason about what explicit permissions may dynamically lead to implicit permissions through complex interactions and uncover 593 implicit permissions derived from explicit permissions. We use the implicit permission references as insights to identify potential risks of CNCF projects and applications provided by the top four cloud vendors. With responsible disclosure, we obtain five new CVEs, six acknowledgments of cloud vendors, and a bounty awarded by Google. These acknowledgments underlie the practical impact of our attack.
The ARTMAN workshop aims to bring together academic researchers and industry practitioners from diverse domains, primarily security & privacy and machine learning, but also various application fields, to collaboratively explore and discuss resilient and trustworthy machine learning-powered applications and systems. This workshop focuses on AI/ML application domains and welcomes contributions on both foundational and applied aspects of ML across various industries, including transportation, aerospace, healthcare, energy, and finance, among others, showcasing AI-driven advances in performance and efficiency. This workshop also seeks contributions on the application of reliable and secure AI/ML algorithms, especially knowledge-informed approaches, to improve resilience and trust, particularly in human-machine partnerships and interactions within such scenarios.
Deep learning techniques are increasingly being incorporated into NIDS. However, the evaluation of such deep learning models often assumes static data distributions and overlooks the effects of randomness and environmental variation. As a result, the reported performance may not reflect the NIDS behaviour during real-world deployment. This paper investigates the impact of stochastic and environmental factors on the evaluation of deep learning models for NIDS, with a focus on shift-aware models that detect and adapt to data shift, representing state-of-the-art systems for long-term deployment. We examine two baselines under controlled variations to analyse the impact of each factor on the reproducibility and fairness of the results, revealing that the F-1 score can vary largely due to these, even minor, variations. All of the explored factors affect the reproducibility of the results, and some can significantly skew performance. Based on our findings, we provide practical recommendations to support reproducible and fair evaluations of deep learning-based NIDS systems.
The use of biometric-based security protocols is on the steep rise. As biometrics become more popular, we witness more attacks. For example, recent BrutePrint/InfinityGauntlet attacks showed how to brute-force fingerprints stored on an Android phone in about 40 minutes. The attacks are possible because biometrics, like passwords, do not have high entropy. But unlike passwords, brute-force attacks are much more damaging for biometrics, because one cannot easily change biometrics in case of compromise. In this work, we propose a novel provably secure Brute-Force Resistant Biometrics (BFRB) protocol for biometric-based authentication and key reconstruction that protects against brute-force attacks even when the server storing biometric-related data is compromised. Our protocol utilizes a verifiable partially oblivious pseudorandom function, an authenticated encryption scheme, a pseudorandom function, and a hash. We formally define security for a BFRB protocol and reduce the security of our protocol to the security of the building blocks. We implement the protocol and study its performance for the ND-0405 iris dataset.
We present SlicedPIR, a distributed Private Information Retrieval (PIR) protocol. SlicedPIR efficiently alleviates the server's compute bottleneck by offloading its load across multiple untrusted client machines. In contrast to prior work, SlicedPIR induces only a modest network overhead when the server offloads its work. It achieves those communication savings by exploiting the polynomial encoding of homomorphic encryption schemes typically used in PIR protocols. This encoding lets the server make novel use of the Number Theoretic Transform (NTT) to distribute points on the polynomials as "slices" of its data rather than the polynomials themselves. Using NTT allows the clients to process recursive PIR queries on their slices and return a succinct result to the server. The server efficiently verifies the clients' results by leveraging the Schwartz-Zippel lemma, which we adapt to the PIR use case. We show how to integrate SlicedPIR into a private messaging system, where clients write messages to the server's database and then use PIR to secretly query for messages from their friends. We implement a prototype of SlicedPIR and run experiments to show that it scales well with the number of clients and database size. Concretely, SlicedPIR achieves better performance and cuts network usage by over 95% compared to the state-of-the-art.
Extended Reality (XR) headsets are increasingly serving as repositories for substantial volumes of sensitive data and gateways to web applications. This transition highlights the need for convenient and secure user authentication solutions. Traditional password/PIN-based schemes are ill-suited to the XR's gesture- and voice-based interfaces and are prone to shoulder-surfing attacks. Some recent XR systems incorporate two-factor authentication, but it requires additional operations on a second device (e.g., a smartphone or wearable). In this work, we introduce the first effortless and inbuilt XR user authentication system by leveraging the harmonics of vibrations excited by users' vital signs. The system is transparent to users (no efforts during enrollment and authentication) and requires no additional hardware. The key idea is that vital signs (i.e., breathing and heart beating) naturally generate low-frequency mechanical vibrations, causing human skull to vibrate and produces harmonic signals. When the harmonics pass the human head, they carry rich biometrics associated with the wearer's skull structure and soft tissues, which can be captured by the XR motion sensors. Instead of directly utilizing the vibrations, we extract more reliable biometrics from the ratios among different harmonic frequencies, which capture wearers' unique head and facial attenuation properties and are non-volatile when the periodicity and amplitude of vital signs fluctuate. We further design an adaptive filter to mitigate the body motion distortions in common XR interactions. By adopting advanced deep learning models with the attention mechanism, our system realizes effective and robust authentication across XR scenarios. Evaluations across 10 months, with 52 users and two popular XR headsets, show that our system can accurately authenticate users with over 95% true positive rates and rejects unauthorized users with over 98% true negative rates under various XR scenarios, with biometrics remaining consistent over long-term periods.
Industrial Control System (ICS) and its software touches every aspect of the critical infrastructure used by our industry, academia, and government. Back in the days, these systems and software were not designed with security in mind. With the ever expanding interconnectivity of ICS environments and new threats, practitioners are stuck on a patchwork of security. While certain proprietary ICS software manufacturers have started to provide security solutions, free and open source ICS software is often less known. The goal of the workshop is twofold: we want to collect ideas on redesigning (parts of) the ICS ecosystem so that security is built-in by design; we also invite contributions on designing, incorporating, and maintaining secure open-source ICS software. The RICCS '25 complete workshop proceedings can be found at: https://dl.acm.org/citation.cfm?id=3733823
Homomorphic encryption (HE) is a foundational technology in privacy-enhancing cryptography, enabling computation over encrypted data. Recently, generalized HE primitives designed for multi-party applications, such as multi-party HE (MPHE), have garnered significant research interest. While constructing secure multi-party protocols from MPHE in the semi-honest model is straightforward, achieving malicious security remains challenging as it requires zero-knowledge arguments of knowledge (ZKAoKs) for MPHE ciphertexts and public keys. In this work, we design practical ZKAoKs for MPHE that validate the well-formedness of public keys and ciphertexts. Specifically, we develop our ZKAoKs within the polynomial interactive oracle proof (PIOP) framework. To achieve this, we introduce novel optimization techniques that seamlessly integrate constraints for MPHE into the PIOP framework, enabling the design of PIOPs for validating all types of MPHE public keys, including relinearization and automorphism keys. To the best of our knowledge, our construction is the first ZKAoK for MPHE that validates automorphism keys. We instantiate our PIOP using a lattice-based polynomial commitment scheme (PCS). When compared with the previous state-of-the-art construction, PELTA (ACM CCS 2023), our implementation achieves a 5.4x reduction in proof size, a 111x speed-up in proof generation, and a 768x improvement in verification time for validating the encryption key. In addition to the encryption key, we provide benchmark results for all types of ZKAoKs required for MPHE, presenting the first concrete performance results in compiling passively secure MPHE-based protocols into maliciously secure ones.
Mobile super apps, which bundle multiple mini-apps into a single platform, have become central to the consumer-facing digital ecosystem. Services such as WeChat, Alipay, Grab, and TikTok integrate payments, messaging, commerce, and entertainment, while at the same time collecting and processing large volumes of sensitive personal data. This concentration of functionality creates unprecedented opportunities to businesses and online services but also raises significant security and privacy risks. Meanwhile, a growing trend is the integration of large language models (LLMs) into mobile apps, transforming them into LLM-driven agentic systems. These systems are capable of orchestrating mini-apps and other mobile apps, interacting with external services, and carrying out privileged tasks on behalf of users. While this enables powerful new applications, it also expands the attack surface and introduces new forms of data exposure, misuse of privileges, and adversarial manipulation. The workshop seeks contributions including but not limited secure architectural design, permission frameworks, threat modeling, privacy-preserving methods, and case studies of real-world deployments in order to build a foundation for safer and more trustworthy LLM-empowered super apps and more generally mobile apps.
Deep learning approaches have achieved remarkable performance in malware classification and detection. However, their success relies on the availability of large, accurately labeled datasets: a critical yet challenging requirement in the malware domain. In practice, most malware datasets are automatically labeled using outputs from antivirus engines, a process that often introduces significant label noise. Such imperfections can severely degrade the performance and generalizability of deep learning models. To address this challenge, we introduce SLB, a framework designed to robustly train deep learning-based malware systems while simultaneously refining dataset labels. SLB begins by partitioning the dataset into two subsets: a clean set containing samples with reliable labels, and a noisy set with samples that may be mislabeled, to which pseudo labels are assigned. As training progresses, SLB continuously monitors the model's predictions to dynamically update both sets. Specifically, samples in the noisy set that consistently receive predictions aligning with their (observed or pseudo) labels are promoted to the clean set, whereas samples in the clean set that exhibit unstable predictions are reclassified as noisy. This iterative process not only enhances model performance but also progressively corrects labeling errors. We evaluated SLB on multiple security datasets with both synthetic and real-world label noise across various deep learning architectures and ML algorithms. Experimental results show that SLB significantly improves malware detection performance and reduces overall noise. For example, on the Android binary dataset with 25% injected label noise, SLB reduced the noise to below 1.5% while increasing the macro F1 score from 74.51% to 96.03% and the accuracy score from 87.66% to 98.68%.
Software developers face risks of leaking their software secrets, such as API keys or passwords, which can result in significant harm. Secret management tools (SMTs), such as HashiCorp Vault Secrets or Infisical, are highly recommended by industry, academia, and security guidelines to manage secrets securely. SMTs are designed to help developers secure their secrets in a central location, yet secrets leaks are still commonplace, and developers report difficulty in learning how to setup and use SMTs. While SMTs typically come with publicly available help resources (e.g., tool documentation and interfaces), it is unclear if these actually help developers learn to effectively use SMTs. Without usable help resources that onboards developers, quick adoption and effective use of SMTs may be unrealistic. In a qualitative two-step study, we observed 21 new users in person while they used SMTs to perform two secret management tasks: secret storage and access, then secret injection. We interviewed participants after each task to identify their challenges and experiences using SMTs, with the assistance of help resources. While our study sample is narrow, it serves as a reasonable proxy for new developers who are likely to adopt SMTs early in their careers. We found that even in a laboratory setting where new users found tool functionality and interface flexibility helpful, they still experienced increased difficulty to effectively use SMTs to securely remediate a hard-coded secret when they felt tool documentation was insufficient. Insufficient tool documentation motivated participants to deviate from official tool documentation to access secondary sources or attempt workaround methods. Specific challenges reported by participants were tool documentation content quality, navigation difficulties with both tool documentation and web interfaces for finding helpful content, and supportive tool features. We explain how these challenges negatively affect participant experiences adopting SMTs, and suggest recommendations on tool documentation and interfaces for SMT developers. If developers cannot simply and quickly manage secrets securely, secret leakage will continue to be commonplace.
The layered architecture of the TCP/IP protocol stack enables protocol layers to be implemented independently and flexibly. However, this layered design introduces potential security risks when shared resources are not properly managed between different layers. This paper investigates a neglected cross-layer shared resource risk, termed SocketFilled, which exploits the insecure usage of the UDP send buffer at the transport layer by the link layer, resulting in the interruption of response packets from the upper application layer. To explore the root causes of cross-layer DoS vulnerabilities resulting from the implementation of the TCP/IP protocol stack, we systematically analyzed the protocol standards of address resolution and reviewed the implementation in mainstream open-source operating systems. Moreover, we conducted a comprehensive experimental evaluation of mainstream operating systems (e.g., Linux and FreeBSD) and UDP services (e.g., DNS and QUIC). The experimental results show that the latest version of Linux and UDP service software (e.g., BIND9, PowerDNS, and Nginx) are affected, causing significant packet loss and even complete service interruption. Then, we estimated the impact range of SocketFilled in the wild and demonstrated that 17.3% of open resolvers, 54.3% of authoritative servers of the Tranco Top 100K domains, and 3.8% of these well-known domains' HTTP/3 servers are potentially affected, including Bing, Amazon, and Shopee, after excluding the influence of cloud servers. We have conducted responsible disclosure by reporting the vulnerability to the Linux community. Our research highlights the effectiveness of cross-layer mechanisms in DoS attacks and calls for heightened attention to the layered complexity of protocol stack implementations within the security community.
Security protocols often involve loops, such as for ratcheting or for manipulating inductively-defined data structures. However, the automated analysis of security protocols has struggled to keep up with these features. The state-of-the-art often necessitates working with abstractions of such data structures or relies heavily on auxiliary, user-defined lemmas. In this work, we advance the state-of-the-art in symbolic protocol verification by adapting cyclic induction proof systems to the security protocol domain. We introduce reasoning rules for the Tamarin prover for cyclic proofs, enabling new, compact proofs, and we prove their soundness. Moreover, we implement new, simple, and effective proof search strategies that leverage these rules. With these additions, Tamarin can prove many lemmas that previously required, often complex, auxiliary lemmas. We showcase our approach on fourteen case studies, ranging from toy examples to a detailed model of the Signal protocol. Our work opens an exciting new research area where automatic induction helps scale security protocol verification, as we provide a fundamentally new and general induction mechanism.
Garbled Circuits (GC) is a foundational primitive for secure two-party computation (2PC). Garbled Private Information Retrieval (GPIR) is a GC technique for looking up a public array or database (DB) on a private index unknown to either player. GPIR immediately implies GC evaluation of functions implemented as a publicly known look-up table (LUT). GPIR is costly: it can be obtained by a linear scan, adapting Garbled RAM, stacking GC branches implementing access to table elements, and, most recently, from GC Look-up Table logrow (Heath et al., Eurocrypt 2024). For a N-row DB lookup of m-bit rows, logrow's computation is approximate to O(Nm kappa), and communication is O(m(log N . kappa + N)). logrow thus can be effectively used on tables of sizes up to approximate to 2(15). We propose Toss, a new efficient GPIR with dramatically reduced bandwidth consumption (a scarce resource in MPC!), both asymptotically and concretely. Our communication cost is O(root Nm root kappa) with a small constant, sublinear both in N and security parameter kappa. Our computation cost is O(Nm kappa + (root N/kappa m + N)c(kappa)), where c(kappa) is the computational cost of hash evaluation. This computation cost is about the same as or slightly lower than logrow's. In concrete terms, for a 2(20)-row LUT of 8-bit items, we improve over logrow by factor >31x in communication. On a laptop over 100Mbps channel, the throughput increases from approximate to 10.6 lookup/s to approximate to 81 lookup/s, achieving >7.5x improvement. For 10Mbps channel, Toss's throughput is >28x better. Communication improvement grows with N: e.g., for N = 2(25), m = 32, improvement is > 512x. Toss builds on stacked garbling (SGC) and logrow with multiple nuanced low-level optimizations, requiring reworking of their internals and interfaces. We emphasize that constructing GPIR directly from SGC incurs logarithmic overhead in computation, which actually decreases throughput in typical "laptop + LAN" testbeds. We implement our construction and report on its performance.
