Despite considerable efforts on making them robust, real-world AI-based systems remain vulnerable to decision based attacks, as definitive proofs of their operational robustness have so far proven intractable. Canonical robustness evaluation relies on adaptive attacks, which leverage complete knowledge of the defense and are tailored to bypass it. This work broadens the notion of adaptivity, which we employ to enhance both attacks and defenses, showing how they can benefit from mutual learning through interaction. We introduce a framework for adaptively optimizing black-box attacks and defenses under the competitive game they form. To assess robustness reliably, it is essential to evaluate against realistic and worst-case attacks. We thus enhance attacks and their evasive arsenal together using reinforcement learning (RL), apply the same principle to defenses, and evaluate them first independently and then jointly under a multi-agent perspective. We find that active defenses, those that dynamically control system responses, are an essential complement to model hardening against decision-based attacks; that these defenses can be circumvented by adaptive attacks, something that elicits defenses being adaptive too. Our findings, supported by an extensive theoretical and empirical investigation, confirm that adaptive adversaries pose a serious threat to black-box AI-based systems, rekindling the proverbial arms race. Notably, our approach outperforms the state-of-the-art black-box attacks and defenses, while bringing them together to render effective insights into the robustness of real-world deployed ML-based systems.

This work investigates the application of federated learning for detecting web bots in edge computing settings. The key challenge lies in developing machine learning models that are not only accurate but also robust against evasion attacks, where adversarial actors attempt to bypass detection. Addition-ally, the models must be privacy-preserving to protect sensitive information, ensuring that confidential data is not exposed to third parties during the learning process. By leveraging federated learning, the proposed approach enables collaborative model training across distributed edge nodes without sharing raw data, maintaining user privacy while enhancing detection capabilities against sophisticated web bot attacks.

ML-based malware detection on dynamic analysis reports is vulnerable to both evasion and spurious correlations. In this work, we investigate a specific ML architecture employed in the pipeline of a widely-known commercial antivirus, with the goal to harden it against adversarial malware. Adversarial training, the most reliable defensive technique that can confer empirical robustness, is not applicable out of the box in this domain, for the principal reason that gradient-based perturbations rarely map back to feasible problem-space programs. We introduce a novel Reinforcement Learning approach for constructing adversarial examples, a constituent part of adversarially training a model against evasion. Our approach comes with multiple advantages. It performs modifications that are feasible in the problem-space, and only those; thus it circumvents the inverse mapping problem. It also makes it possible to provide theoretical guarantees on the robustness of the model against a well-defined set of adversarial capabilities. Our empirical exploration validates our theoretical insights, where we can consistently reach 0% Attack Success Rate after a few adversarial retraining iterations.

In this paper, we introduce a new decentralized access control framework with transitive delegation capabilities that tackles the performance and scalability limitations of the existing state-of-the-art solutions. In order to accomplish this, the proposed solution is anchored in the self-sovereign identity (SSI) paradigm, which embodies a distributed identity management system. By adopting this paradigm, we obviate slow cryptographic premises such as identity-based encryption (IBE) that were used in prior work. Furthermore, we enhance the existing verifiable credentials (VCs) from this paradigm by introducing our own decentralized permission objects to support the concept of transitive delegations. This concept allows delegates to further delegate their access to resources with the same or fewer privileges to other entities within the framework. This renders our solution suitable for diverse scenarios, including applications in decentralized building access management. To the best of our knowledge, we are the first to introduce the concept of transitive delegations in this paradigm. Finally, our performance experiments show a performance enhancement of three orders of magnitude compared to the prevailing state-of-the-art solutions.

Biometric authentication systems have traditionally relied on the False Match Rate (FMR) to evaluate security against impersonation threats. However, this metric alone is insufficient for assessing vulnerabilities to statistical attacks because it cannot account for the non-uniformity of mismatches and atypical inputs that adversaries may manipulate. To address this issue, we propose a new evaluation framework that overcomes these limitations. The framework includes an estimate of the effective key space of biometrics and metrics that consider non-uniformity in the biometric embedding space. Our findings demonstrate that our framework provides a nuanced understanding of biometric security. Moreover, optimizing for the proposed metric leads to better security against statistical attacks than optimizing the FMR. Furthermore, the framework provides a comparative security analysis with traditional methods like passwords and PIN codes. It also quantifies the impact on security when adversaries partially know their victims, e.g., demographics.

Biometric authentication has made great strides throughout the years thanks to better hardware and software support. However, attackers are unrelenting in finding new ways to spoof a subject, hereby breaking existing presentation attack detection schemes. Similar to combining multiple authentication factors, a combination of liveness detection defenses is expected to strengthen security against spoofing attacks. The problem that we address is that many defenses have only been evaluated in isolation or in ideal circumstances. In this work, we demonstrate how different liveness components for face authentication can interfere with one another, thereby jeopardizing security. Furthermore, contextual and environmental influences can endanger their robustness. In this work, we propose a security framework for biometric authentication that supports adaptive liveness detection by reasoning upon undesired interactions between defenses, the impact of new attacks, and the context in which they emerge. We validate the flexibility of our framework to account for both historic and novel interplays between attacks and defenses. Our experiments show that our framework effectively accounts for undesired interactions while only incurring a limited and acceptable performance overhead.

Biometric recognition systems typically use a fixed threshold to differentiate between legitimate users and imposters. Yet, this method can be problematic due to differences in individual user performance, whereas some users are more easily recognizable than others. Furthermore, fixed thresholds require extensive tuning on a large test set a priori to determine an optimal threshold value. Adaptive thresholds address these shortcomings by adjusting threshold values based on population characteristics. However, our research demonstrates that adaptive thresholds suffer from a significant weakness as they inadvertently increase the attack surface against face recognition systems. We do so by introducing a novel attack, the doppelgänger attack, where a malicious actor inserts adversarial examples that mimic legitimate users and increase the false rejection rate for these legitimate users by 70%.

This paper presents and systematically evaluates the first masterkey attack against free-text keystroke dynamics. A masterkey is a typing sequence that matches, hence successfully impersonates, a large part of the population. Therefore, masterkeys are effective tools for an adversary who aims to impersonate someone without knowledge of their typing behavior. On top of the attack itself, we present a new unifying evaluation framework for masterkey attacks that allow for the comparison with knowledge-based authentication factors. In other words, we unify the evaluation of password security with that of masterkey attacks and demonstrate that typing biometrics is approximately 20 times less secure than passwords and approximately two times less secure than a 4-digit pin. Lastly, we study the effect of demographics on typing biometrics, which, among others, provides novel insights into the effect of being a well-versed typist on security.

Sharing cyber threat intelligence helps organizations analyze and protect against a growing number and sophistication of security threats. However, organizations are reluctant to share their locally collected cyber threat intelligence with third parties because of the the risk of incidentally disclosing sensitive business data or personally identifiable information, and the subsequent reputational harm or even financial repercussions imposed by the GDPR. To address the different confidentiality needs of threat intelligence producers and consumers, we present and evaluate a practical polyglot solution for privacy-preserving sharing and analysis of confidential or private information, and this on top of a contemporary cyber threat intelligence platform. Additionally, we investigate the security impact and computational overhead of these techniques to analyze correlations between threat events in a privacy-preserving manner and across sharing organizations.

Since their inception, Captchas have been widely used as reverse Turing tests for combating bot proliferation on the web. This has resulted in an arms race between bot developers that automate Captcha solvers and Captcha services that adjust the challenges accordingly or come up with new ones altogether. Ultimately, older generations could be bypassed consistently, and thus in the third version of reCAPTCHA, Google offers zero user friction. The intent in the new system is not only to avoid interrupting user experience but to also obfuscate the nature of the challenge itself, being much less prominent than a text or image recognition task. We introduce a methodology that learns through interaction how to evade detection, while collecting and analyzing reCAPTCHA v3 scores over fifteen months and various web environments. With reinforcement learning as the backbone, we build models that can simulate human-like web browsing behaviour by using the returned score as an informative signal. Our study exposes an important vulnerability: while the score is influenced by a multitude of undisclosed factors, it is easily accessible and it enables adversaries to learn and perfect evasive models. Notably, we demonstrate that our automation models, which integrate general web browsing capabilities, transfer between websites with an evasion rate up to 99.6%.

The OAuth 2.0 protocol is a popular and widely adopted authorization protocol. It has been proven secure in a comprehensive formal security analysis, yet new vulnerabilities continue to appear in popular OAuth implementations.

We propose, design, and evaluate PIVOT, a privacy-enhancing and effective contact tracing solution that aims to strike a balance between utility and privacy: one that does not collect sensitive information yet allowing effective tracing and notifying the close contacts of diagnosed users. PIVOT requires a considerably low degree of trust in the entities involved compared to centralized alternatives while retaining the necessary utility. To protect users’ privacy, it uses local proximity tracing based on broadcasting and recording constantly changing anonymous public keys via short-range communication. These public keys are used to establish a shared secret key between two people in close contact. The three keys (i.e., the two public keys and the established shared key) are then used to generate two unique per-user-per-contact hashes: one for infection registration and one for exposure score query. These hashes are never revealed to the public. To improve utility, user exposure score computation is performed centrally, which provides health authorities with minimal, yet insightful and actionable data. Data minimization is achieved by the use of per-user-per-contact hashes and by enforcing role separation: the health authority act as a mixing node, while the matching between reported and queried hashes is outsourced to a third entity, an independent matching service (MS). This separation ensures that out-of-scope information, such as users’ social interactions, is hidden from the health authorities, whereas the MS does not learn users’ sensitive information. To sustain our claims, we conduct a practical evaluation that encompasses anonymity guarantees and energy requirements.

As biometric templates consist of highly correlated features, the real security level offered by biometric authentication systems remains an open research question. In this work we provide new approximations and a lower bound of the security offered by fuzzy commitment schemes. Fuzzy commitment cryptosystems and in general biometric template protection schemes play an important role in allowing for remote storage and processing of biometric data, as they mitigate the threat of biometric template leakage. The use of such schemes would alleviate some of the usability constraints imposed by the state-of-practice local use of biometrics. As such we conduct an in-depth security analysis for IMU based gait authentication systems, where we evaluate the effectiveness of attacks within the scope of two well-defined threat models that target both unprotected and protected systems. A pivotal enabler of our analysis is the development of nine different approaches to gait authentication, which allows us to perform intramodal fusion on these distinct, yet highly correlated biometric templates, and to protect them with a strengthened fuzzy commitment scheme. Our analysis clearly demonstrates the high correlation between the different biometric templates, which, among others, further showcases the threat of biometric template leakage. Furthermore, as our analysis incorporates a threat model that assumes biometric template leakage, it provides metrics for the security provided by the biometric modality itself.

Computer security incident response teams typically rely on threat intelligence platforms for information about sightings of cyber threat events and indicators of compromise. Other security building blocks, such as Network Intrusion Detection Systems, can leverage the information to prevent malicious adversaries from spreading malware across critical infrastructures. The effectiveness of threat intelligence platforms heavily depends on the willingness to share among organizations and the responsible use of sensitive information that may potentially harm the reputation of the reporting organization. The challenge that we address is the lack of trust in the source providing the threat intelligence and the information itself. We enhance our security framework TATIS—offering fine-grained protection for threat intelligence platform APIs—with distributed ledger capabilities to enable reliable and trustworthy threat intelligence sharing with the ability to audit the provenance of threat intelligence. We have implemented and evaluated the feasibility of our distributed framework on top of the Malware Information Sharing Platform (MISP) solution, and we evaluate the performance impact using real-world open-source threat intelligence feeds.

Context-aware systems in intelligent environments digest large amounts of data and personal information to gain situational awareness as a way to assist individuals with their daily activities, enhance their experiences and adapt to their needs and intention, whenever and wherever they are. Large amounts of data drive these environments, motivating the adoption of big data and cloud technologies. A similar digital transformation is taking place in the Factory of the Future and Industry 4.0, two paradigms on creating smart products through smart processes and procedures. As the attack surface for security and privacy threats grows, it is no surprise that new regulations and directives will be put in place to protect the privacy of individuals. In this paper, we discuss the foundational principles of Privacy by Design and key obligations of the upcoming EU General Data Protection Regulation (GDPR), and highlight how they impact the design and development of context-aware intelligent environments. We provide technical guidelines for better compliance with these regulatory frameworks.

The exponential data growth in intelligent environments fuelled by the Internet of Things is not only a major push behind distributed programming frameworks for big data, it also magnifies security and privacy concerns about unauthorized access to data. The huge diversity and the streaming nature of data raises the demand for new enabling technologies for scalable access control that can deal with the growing velocity, volume and variety of volatile data. This paper presents SparkXS, an attribute-based access control solution with the ability to define access control policies on streaming latent data, i.e. hidden information made explicit through data analytics, such as aggregation, transformation and filtering. Experimental results show that SparkXS can enforce access control in a horizontally scalable way with minimal performance overheads.

Nowadays, most mobile applications rely on device screen locking mechanisms for ensuring practical security, which expects the users to explicitly authenticate with a PIN or biometric irrespective of the perceived threats. Owing to this usability issues, many avoid using device locks potentially compromising the security. To overcome the limitations of this binary approach, we present an implicit authentication framework called PRISM (Policy-driven Risk-based Implicit locking for improving the Security of Mobile end-user devices). It provides risk based authentication by detecting anomalies in the usual behavior patterns of the users which include their expected locations, activities and application usage. Its device unlocking decisions are driven by policies that are defined either automatically by mining sensor data or manually by the end-users. Our experiments show that PRISM is able to discover useful behavior patterns efficiently even with limited data. The number of required explicit authentications is significantly reduced while assuring the preferred security for everyday scenarios.

In the Internet of Things, heterogeneous and distributed streams of sensor events is a driver for context-aware behavior in intelligent environments. However, processing the event data usually cross-cuts the business logic of IoT applications and offering such reusable functionality as a service towards a variety of customers with different needs is often faced with scalability concerns. We present SAMURAI, a multi-tenant streaming context architecture that integrates and exposes well-known components for complex event processing, machine learning, knowledge representation, NoSQL persistence and in-memory data grids. SAMURAI pursues a twofold approach to achieve scalability: (1) distributed deployment with horizontal scalability, (2) shared resources through multi-tenancy. For the scenario used in the experimental evaluation of our architecture, the results show little overhead to support multi-tenancy, with near-linear scalability and flexible elasticity for deployment schemes with data partitioning per tenant.

The intelligent home environment is a well-established example of the Ambient Intelligence application domain. A variety of sensors and actuators can be used to have the home environment adapt towards changing circumstances and user preferences. However, the complexity of how these intelligent home automation systems operate is often beyond the comprehension of non-technical users, and adding new technology to an existing infrastructure is often a burden. In this paper, we present a home automation framework designed based on smart widgets with a model driven methodology that raises the level of abstraction to configure home automation equipment. It aims to simplify user-level home automation management by mapping high-level home automation concepts onto a low-level composition and configuration of the automation building blocks with a reverse mapping to simplify the integration of new equipment into existing home automation systems. Experiments have shown that the mappings we proposed are sufficient to represent household appliances to the end user in a simple way and that new mappings can easily be added to our framework.

In the near future vehicles will be equipped with embedded communication capabilities in order to perform context-sensitive tasks such as traffic flow control and incident avoidance. In large scale vehicular networks it is of vital importance that the information exchanged between the vehicles meets a certain level of quality so that they can make informed automated decisions. In this paper we define Quality of Context (QoC) and Peer Reputation (PR) for nodes in vehicular networks and propose ways to apply them for efficient communication. Thus, we provide a two-fold solution in which on the one hand we focus on the quality of the information and on the other hand we aim at determining the reputation of the nodes involved in the communication. This helps to eliminate the use of erroneous, ambiguous and imprecise information provided by unknown entities. Our simulated experiments show that our mechanism significantly reduces network traffic usage, eases out the decision making process and guarantees a minimum level of quality.

The applicability of pervasive and mobile computing in the health care sector is beyond dispute. This paper explores the use of the mobile phone as a tool for personalized health care assistance for individuals diagnosed with diabetes. By monitoring user location and activity on the mobile phone, recognizing past behavior and augmenting the logging of blood glucose levels with context data, our prototype application assists with taking well-informed decisions on daily drug dosage to achieve and maintain stable blood glucose levels. Evaluation of the prototype application indicate that a brief training of the application suffices to capture patterns in the user's relevant context that simplify glucose level trends analysis. We describe some of the details of the user study and the prototype application, and conclude with plans to investigate context-driven activity prediction to further improve the decision support for the user.