Evaluations of advice on online security, safety, and privacy given to end-users focus mostly on adults whereas it is unclear how these evaluations apply to advice given to children. To address this, we gathered and analyzed such advice with the help of handcrafted risk scenarios through an in-person study. Our scenarios are designed to cover multiple categories of online risks that threaten the security, safety, and privacy of children. This allows us to explore a broad spectrum of advice instead of focusing deeply on single technologies. Each scenario describes a child in a risky scenario and asks a study participant to provide advice to the child. We used these scenarios to collect advice from 287 children (aged 10–13), 195 parents, and 26 teachers. Our qualitative analysis of the resulting pieces of advice shows that participants generally recognize risks and provide useful preventive advice in scenarios in which an incident has not yet occurred. However, participants often misjudge the severity of digital threats and do not provide adequate advice in post-incident scenarios. Based on these findings, we propose avenues for further research and discussion, such as enhancing children’s agency in risk management through targeted education, developing practical response toolkits, and fostering community support among families.
Trigger-based watermarking aims to protect the intellectual property of machine learning models by embedding abnormal behavior that is activated only on specific trigger inputs, while preserving performance on standard test data. Although widely studied, current approaches to trigger-based watermarking remain largely informal and empirical. In this work, we introduce a formal framework for reasoning about the security of trigger-based watermarking, and we propose rigorous security definitions to evaluate existing schemes. Our analysis reveals that many schemes rely on underspecified parameters and flawed design assumptions. Notably, we demonstrate that different security goals in watermarking can be in tension with one another, and achieving a balance among them requires careful design. Our principled analysis offers a way to resolve this tension in practice. Experiments on existing schemes and public datasets corroborate the relevance of our theoretical findings.
Federated Recommender Systems (FRSs) enable on-device personalization while preserving user privacy, yet the gradients shared during model training can still leak sensitive information. Local differential privacy (LDP) offers a formal defense, but classic mechanisms—randomized response and Laplace perturbation—inject substantial noise into the high-dimensional gradient vectors of latent-factor models, severely degrading recommendation accuracy. In this paper, we systematically evaluate these traditional LDP strategies in federated matrix factorization under both explicit (MovieLens 100K and 1M) and implicit (Steam, LastFM) feedback benchmarks. We then propose a novel directional noise mechanism based on the von Mises–Fisher distribution, which preserves gradient magnitudes while randomizing directions. Through extensive experiments across a wide range of privacy budgets, we show that our vMF mechanism consistently outperforms Laplace and binary randomized-response, recovering up to more than 100 % of the non-private utility at ε = 1 and delivering higher ranking accuracy under moderate privacy levels (ε ≤ 2). We further demonstrate that directional perturbation reduces variance and remains robust under partial client participation. Our findings highlight the importance of data geometry in privacy–utility trade-offs and provide practical guidance for deploying LDP in high-dimensional federated recommendation settings.
Account ecosystems comprise a set of resources, including online accounts, devices, and credentials, that jointly impact a user’s or organisation’s security. Prior work shows that users frequently overlook subtle dependencies within these ecosystems, leading to security breaches that are difficult to detect using analysis techniques focused on isolated components.We introduce resource access graphs (RAGs), a formal model for the automated analysis of dynamic account ecosystems. RAGs extend previous work on account access graphs (AAGs) with partially ordered labels for vertices and edges, enabling both general and resource-specific actions to be modelled. To capture the dynamic nature of account ecosystems, we define a transition system based on graph rewriting with RAGs as states in which both user actions and adversarial capabilities are modelled as graph transformation rules. Security properties are specified in computation tree logic (CTL) allowing analysis using existing graph-rewriting based model checkers, such as GROOVE.We demonstrate our approach on a simple Apple account ecosystem with the recent security feature Stolen Device Protection (SDP). Our analysis reveals limitations in SDP, affecting both the integrity and availability of a user’s Apple Account and we propose and verify an improvement to enhance account integrity.
Researchers, security vendors, and network operators rely on DNS intelligence to protect their network infrastructure. Despite the value of this data, the internet-wide monitoring implementations that collect it remain opaque, leaving researchers without a clear understanding of DNS monitor behaviors, monitor coverage, and their exposure to network adversaries. Further, security professionals have publicly reported that DNS replay within the security ecosystem has disrupted the detection of sophisticated attacks. To address this knowledge gap and shed light on current implementation risks, this paper designs and implements a flexible system, DNS Trap, that enables the first broad study of reactive DNS monitoring across the entire IPv4 space. DNS Trap reveals that monitoring pipelines that capture and resolve domains observed in live traffic respond to unsolicited probes sent from our distributed cloud vantage points to more than 30,000 ASes. These resolutions come from 62,479 distinct resolver IP addresses in 4,139 ASes, demonstrating the sweeping network coverage of certain security entities. At the same time, their eagerness to engage with our unsolicited DNS probes creates a clear risk of exposure. Our outside-probing-in approach mirrors an attacker’s ability to conduct reconnaissance and profile network defenses, opening the door for weaponized confusion, avoidance, or active deception. We conclude by using DNS Trap to take a deeper dive into notable individual monitoring behaviors and showcase how network adversaries can reveal behavior during the reconnaissance phase of their attacks. These case studies inform our discussion of risks to network stakeholders and considerations for safer DNS monitoring implementation.
Most research on image steganalysis focuses on pixel-level manipulations, even though many steganographic tools, and almost all stego-malware, embed data in the file structure of image formats, like JPEG. Such payloads are challenging to detect at scale because legitimate JPEGs differ widely across encoders, compression libraries, and social media processing pipelines. This causes distribution drift that inflates errors unpredictably. Since JPEG files can be viewed as a hierarchical tree of markers, we propose to use structure-aware methods, specifically a neural network for hierarchical data (HMill) and hierarchical tree distance (HTD), and compare them with conventional feature engineering (HashTrick). In all tested scenarios, the structure-aware models, especially HMill, achieved the highest detection accuracy and high robustness with respect to distribution drift.
Field-Programmable Gate Arrays (FPGAs) facilitate a write-compile-debug cycle for digital circuits, resembling modern software development. Crucially, FPGA circuits comprise modules that can be re-used across projects. Many generic modules are available from commercial and open-source vendors as Intellectual Property (IP) cores.However, using third-party IP cores in FPGA designs poses critical supply chain security risks: FPGAs are increasingly being used in security- and safety-critical applications like automotive driver assistance or core networking. Thus, vulnerabilities in any component—homegrown or licensed—can have far-reaching consequences.While susceptibility of FPGA IP to security-critical implementation bugs has been shown in the literature, the prevalence and management of security vulnerabilities in FPGA IP has never been studied quantitatively. In this paper, we present the first measurement study covering 19 representative IP cores with 1250 total known issues over three commercial vendors and one open-source vendor. We classify 284 issues as security critical, predominantly affecting availability (134) and integrity (184). We analyze the survivability of these vulnerabilities and their mitigations. We find that mean-time-to-fix is between 5 and 12 months for commercial vendors, with outliers of up to 70 months, and no available mitigations for certain vulnerabilities. Specifically, 14 issues have fixes that support fewer FPGA devices than the vulnerable version, including 11 issues that vendors themselves label as critical but for which no workaround exists. We conclude with concrete recommendations for the industry, based on lessons learned from software vulnerability management.
NIST-standardized Deterministic Random Bit Generators (DRBGs) are widely deployed in various operating systems and real-world cryptosystems. Their security has been analyzed under the assumption that the DRBGs use random seeds that are independent of the entropy source. However, DRBG operations might affect the physical state of the hardware, creating a correlation between the seed and the entropy source. So the seedless model better captures the real environment around a DRBG, while security in this model is not immediate from the security in the (stronger) seeded model. For example, the NIST standard CTR-DRBG, which has been proven to be secure under the seeded assumption, turns out to be vulnerable in the seedless model.Motivated by this observation, we prove the security of the NIST-standardized hash-based DRBGs, specifically HASH-DRBG and HMAC-DRBG, in the seedless robustness model. In particular, by modifying and applying the Multi-Extraction framework of Chung et al. to the security proof, we improve the previous security bound of HASH-DRBG from $O\left( {\min \left\{ {{2^{n/2}},{2^{\lambda /2}},{2^{L/3}}} \right\}} \right)$ in the seeded model to an even better bound $O\left( {\min \left\{ {{2^{n/2}},{2^{\lambda /2}},{2^{L/2}}} \right\}} \right)$ in the seedless model, where n is the output size of the underlying hash function, λ is the min-entropy threshold of the entropy source, and L is the size of the state components in HASHDRBG. Our result implies that HASH-DRBG remains secure even with an internal state of a smaller size. Furthermore, in our security proof, a DRBG is allowed to collect randomness of small entropy multiple times before it returns the output. This assumption is, for example, relevant to the security of Linux, where entropy is gradually accumulated over multiple refresh calls. We prove that NIST hash-based DRBGs also remain secure even with such sources of small entropy.
End-to-end (E2E) verifiable remote electronic voting is often discussed as an alternative to postal voting for voters living abroad. Such systems require voters to verify that their vote is recorded as intended and to complain if this is not the case. The challenge with voters living abroad is to provide them with a usable, reliable online channel to file their complaints. The contributions of our paper are: (1) Proposal of such a reliable mechanism for a well-established voting scheme including the adoption of voting material to support it. (2) Evaluation of the proposal’s effect on the usability (i.e. effectiveness, efficiency, and satisfaction of the voting process) and the manipulation detection rate on the voting scheme. (3) Evaluation of the effectiveness of the complaint filing mechanism. The study comprised two phases in which 150 participants cast their votes over the internet after receiving the voting material by postal mail. The first phase focused on the usability and the second one (being deceptive) simulated a manipulation, examining the manipulation detection rate and the effectiveness of the complaint filing mechanism. Our results show a significant improvement in efficiency, without any negative impact on other usability dimensions or participants’ ability to detect manipulations. Among those who detected the manipulation in the second phase, 73% used the complaint filing mechanism correctly, while the rest would have filed a complaint to the attacker who had manipulated their vote in the first place. We conclude that the proposed reliable complaint channel can and should be integrated into E2E verifiable remote voting systems and outline ways to help more voters use it effectively.
Open-source operating systems now power a wide range of devices, from wearables and smart appliances to vehicles. To meet specific functional, security, privacy, or hardware compatibility requirements, developers often customize these systems. However, such changes often carry a cost: downstream maintainers may overlook critical security patches or inadvertently reintroduce previously mitigated vulnerabilities when integrating new features and legacy drivers. In this study, we present a novel framework designed to efficiently detect vulnerable code in Custom Operating Systems (COSs) through a dual-layer architecture that identifies both insecure code clones and the absence of expected security-critical patches. While applicable to any COS, we evaluate our framework on Android and refer to it as AndroVET. Being one of the world’s largest software platforms, Android exhibits systematic fragmentation in security fixes due to its diverse vendor variants and heterogeneous patching practices, making it an ideal use case scenario. We implemented and open-sourced our framework, and thoroughly evaluated it on 14 real-world Android-based COSs. AndroVET detected 332 vulnerability instances (corresponding to 94 unique CVEs) across 14 COSs, and a GrapheneOS case study confirmed 15 absent patches from that distribution. Experimental results show that our framework achieves higher vulnerability-detection accuracy while reducing analysis time, running on average 4 times faster than existing tools.
Intrusion Detection Systems (IDS) are essential tools for detecting and analyzing malicious system activity. Anomaly-based IDS have gained popularity due to their ability to detect zero-day attacks, unlike signature-based IDS. Although this approach is promising, recent AI-based IDS still suffer from high false positive rates, non-scalability and biases, which limit their practicality in real-world deployments.This paper presents GRAAL, an end-to-end unsupervised graph-based anomaly-based IDS that allows a scalable multi-level detection, achieves a low false positive rate, and provides intuitive and interpretable outputs to assist analysts in threat detection and investigation. GRAAL proposes a method to extract features from heterogeneous provenance graphs, using a combination of structural and attribute embeddings. These vectors are then processed by multiple autoencoder models to detect anomalies at both graph and system entity levels. GRAAL’s models leverage the relationship between the graph and entity levels, sharing knowledge through transfer learning and combining their results. We compare GRAAL against six state-of-the-art IDS on several datasets. These comparisons reveal methodological and reproducible biases in the evaluation of the current IDS and lead us to define best practices. To perform a comprehensive comparison, we mitigate biases still present in these state-of-the-art IDS. These extensive evaluations of GRAAL show that GRAAL outperforms these IDS, with a higher precision.
Unlike static biometrics (e.g., faces and fingerprints), dynamic behavioral biometrics are believed to be more difficult to replicate. This paper investigates the security of behavioral biometrics considering the advancements in robotics and AI, particularly as humanoid robots, like Tesla Optimus, are expected to be mass-produced in the coming years. We find that general robotic arms have already gained the capability to reproduce human hand motion trajectories. However, using robots to replicate a user’s behavioral biometrics for attacks remains under-explored due to two long-standing challenges: 1) how to obtain the user’s complex behavioral biometrics through practical eavesdropping (not just trajectories); 2) how to replicate the user’s behavioral kinematics based on the eavesdropped data using a real robot. This work is the first to comprehensively address the two challenges. We develop the point-wise GAN-based Robot Replay Attack (GANRRA) to demonstrate a practical human behavioral replay attack using a hidden camera and a physical robot. GANRRA utilizes a hidden camera to eavesdrop on the user’s hand motions and employs a generative adversarial network to reconstruct the motion data, addressing the sensor discrepancies between the legitimate sensor and the hidden camera and maximizing the behavioral feature similarities. The reconstructed motion data is converted into velocity commands for a robot to execute point by point, replicating both hand movement trajectories and behavioral biometric features. For experiments, we implement an in-air signature system using two existing hand-tracking systems and fool them using a robotic arm attached with a fake hand. Results show that GANRRA reproduces in-air signatures with a 73.1% success rate. To address such robot-relay threats, a novel defense mechanism based on multi-joint behaviors is proposed.
Following the expansion of IoT systems, spanning from devices to cloud backends, reported IoT CVE vulnerabilities have increased at an alarming pace. Since most IoT attacks exploit known vulnerabilities, understanding known vulnerabilities is vital for defense and security research. In this work, we systematize the prior research on studying IoT vulnerabilities, revealing the absence of consistent IoT definitions, reliable and scalable classification methodologies, and high-quality IoT CVE datasets. To overcome these limitations, we design LLIoT, a novel and LLM-assisted approach that systematically and automatically distinguishes IoT-specific CVEs at large scale, enabling in-depth under-standing of IoT vulnerabilities. First, leveraging the systematization knowledge from the literature, we derive a four-layer IoT ecosystem taxonomy and define classification criteria for distinguishing IoT CVEs. Then, using an expert-validated ground-truth dataset, we demonstrate that LLMs can reliably distinguish IoT from non-IoT CVEs with a high accuracy of 95%, outperforming humans by avoiding cognitive errors and gaps in domain knowledge. Applying LLIoT to CVEs from 2013–2024, we build a dataset of 15,116 IoT-specific vulnerabilities, of which 8,368 are newly classified with respect to previous datasets. Using this dataset, which we share with the research community for further research and reproducibility, we characterize how IoT vulnerabilities differ from traditional IT vulnerabilities. Upon our observation, we provide actionable recommendations for responsible stakeholders.
Bluetooth communication is built upon a client-server architecture, where the client initiates service discovery by using Universally Unique Identifiers (UUIDs) to locate the intended service provided by the server. This mechanism was originally designed for single-purpose devices (e.g., headsets), where services are always provided by the sole device vendor and trusted. However, modern smartphones are shared by many apps where these services are instead offered by untrusted and even malicious third-party apps. This shift from a trusted device to an untrusted app model may open the door for service impersonation attacks, as a malicious app on the smartphone can register services with UUIDs used by other apps.In this paper, we investigate Bluetooth service discovery flaws in modern OSes, and confirm the feasibility of this novel intra-device impersonation attack named UUID Confusion Attack. Specifically, this attack results in real-world impacts such as stealing files from a system-level file-sharing app, injecting keystrokes, and forcing a device to join a hostile Wi-Fi network. These end-to-end exploits affect major Android vendors and endanger a user base of over 500 million. The vulnerabilities have been acknowledged by the affected developers, including manufacturers (OEMs) such as Xiaomi, OPPO, and vivo. To mitigate this security threat, we propose a practical defense for Bluetooth on Android. In particular, this defense is a lightweight modification to the Android OS with minor performance overhead, requiring no changes to existing apps or external devices. It provides immediate protection when detecting UUID collisions while introducing minimal interruptions to users.
Understanding the structure of network protocols is necessary for traffic analysis and security auditing. While many protocols are well documented, custom or proprietary protocols often lack public specifications. In this paper, we present MARISSA, an automated tool for network protocol reverse engineering that infers message formats directly from network traces. MARISSA combines similarity digest-based clustering with multiple sequence alignment to identify field boundaries in unknown traffic. Unlike previous tools, it scales efficiently to large datasets and supports a wide variety of textual and binary protocols. We evaluated MARISSA on nine real-world protocols, where it consistently outperformed existing tools in clustering accuracy and field inference quality. Notably, MARISSA achieves substantial improvement in runtime, reducing it by factors ranging from 2× to 15×. We further illustrate its practical utility using a case study of ARPChat, a custom chat application based on the ARP protocol. MARISSA successfully reconstructed message structures and facilitated interoperability with the original application. Our results show significant improvements in both scalability and inference accuracy, underscoring the tool’s effectiveness in analyzing network protocols.
Fuzzing has become one of the most effective automated bug discovery techniques. Despite extensive research covering all aspects, it is difficult to assess the actual progress fuzzing has made over the years. In this paper, we present a large-scale empirical analysis of fuzzing progress using AFL++, the state-of-the-art fuzzer that continuously integrates research-driven improvements. Using 645,000 CPU-hours of experiments, we comprehensively measure how fuzzing has improved in terms of code coverage and evaluate the impact of various features over the years. Surprisingly, we find a plateau in exploring new program behavior: while some techniques yield isolated performance gains, overall progress in exercising new coverage has largely stalled. Studying whether our observations generalize to LibAFL and Fuzzilli, we find our observations hold across all three fuzzers. To better understand this stagnation, we complement our empirical study with a survey of 405 peer-reviewed fuzzing papers published between 2018 and 2024 at the leading security and software engineering venues. We identify 60 papers that extend AFL/AFL++ and study how feasible the integration into the baseline is, and if the baseline fuzzer adopted it. Surprisingly, we observe little adoption in practice, with irreproducible results, reliance on complex external dependencies, and limited practical benefit as the main barriers. Discussing our analysis results with the AFL++ maintainers, we find a growing disconnect between academic research and real-world adoption, underscoring the need for stronger reproducibility standards and a more realistic benchmarking of proposed improvements.
Threat actors have extensively used cloud services as covert channels for Command & Control (C2) and data exfiltration. In response, best practices for enterprise security recommend blocking unsanctioned cloud services and monitoring egress traffic to detect data exfiltration. In this paper, we demonstrate that these defenses fail in the context of first-party services offered by major Cloud Service Providers (CSPs), such as Microsoft Azure, Amazon Web Services (AWS), and Google Cloud Platform (GCP). To provide their services, CSPs require enterprise clients to create broad network openings to some of their services, which means that blocking access to other tenants’ resources for those services is not possible. Additionally, their tight integration into enterprise IT means large volumes of traffic move from the enterprise to the cloud services, making it difficult to distinguish malicious from legitimate activity.To demonstrate the extent of this threat, we implement practical covert C2 channels through six first-party services across Azure, AWS, and GCP. Our simulated attacks demonstrate successful data exfiltration despite mature enterprise-level security controls. These attacks are both stealthy (i.e., evading OS-level and network-level detection) and effective (i.e., achieving practical exfiltration rates of over 400 Mbps). To complement our technical evaluation, we conducted a focus group study with security professionals. Participants acknowledged that awareness of this threat has recently increased, although it remains largely limited to the security experts inside enterprises. They also agreed that detecting such attacks is extraordinarily difficult. Our findings highlight a systemic blind spot in enterprise security and call for coordinated mitigation efforts between CSPs and developers of third-party enterprise services.
Semantic search over vector databases is a fundamental problem in both theory and practice, with large-scale systems relying on approximate nearest-neighbor (ANN) algorithms to retrieve semantically similar results efficiently. Achieving this capability securely while keeping both data and queries hidden remains a major challenge. Existing secure semantic search systems incur high latency and fail to scale to realistic database sizes.We present BiSON , the first secure nearest-neighbor search protocol capable of supporting billion-scale encrypted vector databases. Even at this scale, BiSON answers queries in mere milliseconds and maintains search accuracy comparable to state-of-the-art insecure ANN algorithms, demonstrating that secure semantic search can be both private and truly high-performance. Compared to Compass, the prior state-of-the-art system, BiSON reduces communication up to 28×, improves end-to-end latency by up to 23.5×, and scales to datasets that are two orders of magnitude larger.A central contribution of BiSON is a new disk-compatible Oblivious RAM (ORAM) architecture that enables seamless scaling to billion-point datasets without compromising latency or privacy. Together, these innovations make BiSON the first practical and scalable solution for secure semantic search at cloud scale.
Sophisticated rootkits targeting the UEFI boot process are used in attacks to infiltrate devices and maintain persistence across reboots. Despite efforts to protect the boot process, challenges such as a bloated UEFI environment, insecure configurations, and firmware vulnerabilities have made these defenses insufficient. Thus, innovative detection and prevention methods are needed.This work introduces Helltrap, a framework that transforms devices into honeypots capable of detecting rootkits that compromise the UEFI boot process. By leveraging firmware update capabilities, Helltrap automatically integrates monitoring systems within UEFI firmware and Baseboard Management Controller software found on server-grade systems. This allows it to detect modifications to UEFI firmware, Option ROM, or bootloaders, preventing rootkit payloads from executing. The payload is then automatically removed and made available for analysis.To demonstrate the need for Helltrap, we collected real-world rootkit payloads that target UEFI firmware or the ESP partition. Our experiments show that these rootkits can evade detection by standard malware detection tools. To further underscore the danger of such malware, we also developed a proof-of-concept rootkit, GuardDown, which compromises even systems protected by Intel Boot Guard or Secure Boot. In contrast, Helltrap successfully detected all tested rootkits before execution.
