IMDEA Networks Institute Digital Repository
Not a member yet
1915 research outputs found
Sort by
Measuring Alexa Skill Privacy Practices across Three Years
Get PDFSmart Voice Assistants are transforming the way users interact with technology. This transformation is mostly fostered by the proliferation of voice-driven applications (called skills) offered by
third-party developers through an online market. We see how the number of skills has rocked in recent years, with the Amazon Alexa skill ecosystem growing from just 135 skills in early 2016 to about
125k skills in early 2021. Along with the growth in skills, there is increasing concern over the risks that third-party skills pose to users’ privacy. In this paper, we perform a systematic and longitudinal measurement study of the Alexa marketplace. We shed light on how this ecosystem evolves using data collected across three years between 2019 and 2021. We demystify developers’ data disclosure practices and present an overview of the third-party ecosystem. We see how the research community continuously contribute to the
market’s sanitation, but the Amazon vetting process still requires significant improvement. We perform a responsible disclosure process reporting 675 skills with privacy issues to both Amazon and
all affected developers, out of which 246 skills suffer from important issues (i.e., broken traceability). We see that 107 out of the 246 (43.5%) skills continue to display broken traceability almost one year after being reported. As a result, the overall state of affairs has improved in the ecosystem over the years. Yet, newly submitted skills and unresolved known issues pose an endemic risk.TRUEpu
Beyond Voronoi: Plain Probabilistic Spatial Coverage Inference from Base Station Deployments
Get PDFMapping information collected at the level of individual base stations onto the geographical space is a required operation for many works relying on mobile network metadata. The common practice is to represent base station coverage as Voronoi cells, and assume that users are uniformly distributed therein. In this paper, we leverage a large-scale dataset of realistic spatial association probabilities to over 5,000 operational base stations, and quantify the substantial problems of such a simplistic mapping approach. To address the limitations of legacy Voronoi representations, we develop VoronoiBoost, a data-driven model that scales Voronoi cells to match the probabilistic distribution of users associated to each base station. VoronoiBoost relies on the same input as traditional Voronoi decompositions, but provides a richer and more accurate rendering of where users are located: hence, it can be readily used by researchers to substantially improve the spatial representation of mobile network metadata. Our experiments demonstrate that VoronoiBoost improves the quality of mapping by 44% on average over standard Voronoi cells. We also showcase the utility of our model in a practical Edge network planning use case, where the information produced by VoronoiBoost drives a deployment up to 28% more accurate than that obtained with Voronoi cells.Comunidad de MadridFALSEpu
On how VoIP attacks foster the malicious call ecosystem
Get PDFSwitched telephone networks are a key and ubiquitous infrastructure. Recent technological advances have integrated modern and inexpensive systems into these networks in order to use the Internet to place calls via Voice over IP (VoIP). The evolution of this technology has also led to an increase in the number and sophistication of the techniques used by criminals to commit fraud. Specifically, with the emergence of VoIP, attackers can now adapt tools commonly used by cybercriminals, such as botnets, to make their attacks more complex and insidious. For example, through bots they can dial multiple numbers automatically, enabling them to target a greater number of victims, and do so more quickly. While recent studies have shed light on how certain parts of this ecosystem work, it is still unclear how attacks on VoIP systems contribute to this type of fraud. This paper presents a novel VoIP honeypot that captures voice interactions, in addition to employing low-level telemetry. With the study of how attackers obtain access to our honeypot and the actions they perform, we present an overview of the most prevalent types of fraud used in this ecosystem, including unique insights into the origin of the attacks and the destination of calls made through our architecture. Finally, we analyze in-depth the actions taken to study the different types of telephony fraud.Ramon y Cajal RYC-2020-029401-ITRUEpu
Classification of Underwater Fish Images and Videos via Very Small Convolutional Neural Networks
Get PDFThe automatic classification of fish species appearing in images and videos from underwater cameras is a challenging task, albeit one with a large potential impact in environment conserva- tion, marine fauna health assessment, and fishing policy. Deep neural network models, such as convolutional neural networks, are a popular solution to image recognition problems. However, such models typically require very large datasets to train millions of model parameters. Because underwater fish image and video datasets are scarce, non-uniform, and often extremely unbalanced, deep neural networks may be inadequately trained, and undergo a much larger risk of overfitting. In this paper, we propose small convolutional neural networks as a practical engineering solution that helps tackle fish image classification. The concept of “small” refers to the number of parameters of the resulting models: smaller models are lighter to run on low-power devices, and drain fewer resources per execution. This is especially relevant for fish recognition systems that run unattended on offshore platforms, often on embedded hardware. Here, established deep neural network models would require too many computational resources. We show that even networks with little more than 12,000 parameters provide an acceptable working degree of accuracy in the classification task (almost 42% for six fish species), even when trained on small and unbalanced datasets. If the fish images come from videos, we augment the data via a low-complexity object tracking algorithm, increasing the accuracy to almost 49% for six fish species. We tested the networks with images obtained from the deployments of an experimental system in the Mediterranean sea, showing a good level of accuracy given the low quality of the dataset.European Union’s Horizon 2020 Research and Innovation ProgrammeTRUEpu
Implementation and Scalability Evaluation of Random Forests for In-Switch Inference
Get PDFWe comparatively evaluate state-of-the-art solutions for in-switch machine learning inference. We demonstrate that random forest (RF) models attain accuracies on par with those of approaches based on neural networks, which are also less amenable to in-switch operation. Next, we implement the top two solutions for in-switch random forest representation using a unified framework that we propose to ensure their fair comparison. We then verify their performance, resource consumption and scalability with respect to the capabilities of production-grade programmable switches.European Union Horizon 2020 research and innovation program under grant agreement no. 101017109 “DAEMON”European Union Horizon 2020 research and innovation program under Marie Skłodowska-Curie grant agreement no. 860239 “BANYAN”FALSEpu
Coded Caching in Land Mobile Satellite Systems
Get PDFWe investigate the performance of coded caching in land mobile-satellite (LMS) systems, where a satellite station with full access to a content library serves K cache-aided land users. The promising gains that coded caching provides for the error-free shared-link Broadcast Channel are known to suffer in some low signal-to-noise ratio (SNR) scenarios. For this reason, we analyze to what extent the coded caching gains are preserved in LMS systems, which are governed by low-to-moderate SNR due to the long propagation distances. We model the satellite-terrestrial channels through the widely adopted Shadowed-Rician fading, and we show that the coded caching gains are partially preserved even in the low-SNR limit due to the existence of line-of-sight (LOS) components, which allows us to double the goodput of LMS communication at low SNR. These results illustrate the potential of coded caching in LMS systems and motivate further research to design adapted and practical coded caching schemes for LMS systems.Comunidad de Madrid - Atracción de talentoTRUEpu
Try Before You Buy: a Practical Data Purchasing Algorithm for Real-World Data Marketplaces
Get PDFData trading is becoming increasingly popular, as evident by the appearance of scores of data marketplaces (DMs) in the last few years satisfying the demand for third-party data. For buyers, however, deciding whether paying the requested price makes sense can only be done after having tested the data on their ML model. In this paper, we propose a method for optimizing data purchasing decisions. We show that if a marketplace provides to potential buyers a measure of the performance of their models on \emph{individual} datasets, then they can select which of them to buy with an efficacy that approximates that of knowing the performance of each possible combination of datasets offered by the DM. We call the resulting algorithm Try Before You Buy (TBYB) and demonstrate over synthetic and real-world datasets how TBYB can lead to near optimal data purchasing with only O(N) instead of O(2^N) information and execution time.EU Horizon 2020TRUEpu
I move U move: V2X-enabled wireless towing
Get PDFAs smart connected vehicles become increasingly common, their ability to provide enhanced services
has improved. One such service is the emergency transport of drivers in medical distress. In this paper,
we show how such a service can be run from the network and discuss the importance of having a human
in the loop in order to expedite driving. We present a Monte-Carlo-based driver assessment system that
the network can use in selecting the most suitable candidate to tow an autonomous vehicle with an
incapacitated driver. We show that this mechanism results in a selection policy that ensures sufficiently
short spacing between the autonomously driven tail/towed vehicle and the human driven lead/towing
vehicle ensuring that no other vehicles get in the way to disrupt the service.TRUEpu
Automated Identification of Network Anomalies and Their Causes with Interpretable Machine Learning: the CIAN Methodology and TTrees Implementation
Get PDFLeveraging machine learning (ML) for the detection of network problems dates back to handling call-dropping issues in telephony. However, troubleshooting cellular networks is still a manual task, assigned to experts who monitor the network around the clock. To help in this task we present CIAN (from Causality Inference of Anomalies in Networks), a practical and interpretable ML methodology, which we implement in the form of a software tool named TTrees (from Troubleshooting Trees). We have designed CIAN to automate the identification of the causes of performance anomalies in cellular networks. Our methodology is unsupervised and combines multiple ML algorithms (e.g., decision trees and clustering) and Kolmogorov complexity-inspired data analysis tools that we have developed for this work. CIAN can be used with small volumes of data and is quick at training.
Our experiments use diverse data sets obtained from measurements in operational commercial mobile networks. They show that the TTrees implementation of CIAN can automatically identify and accurately classify network anomalies – e.g., cases for which a network low performance is not apparently justified by operational conditions – training with just a few hundreds of data samples. The resulting information hence enables precise troubleshooting actions. In particular, we showcase how TTrees can be flexibly used to monitor the performance of TCP and QUIC protocols when they are adopted to serve mobile users.Comunidad de MadridSpanish State Research Agency (AEI)TRUEpu
“And all the pieces matter...” Hybrid Testing Methods for Android App’s Privacy Analysis
Get PDFSmartphones have become inherent to the every day life of billions of people worldwide, and they are used to perform activities such as gaming, interacting with our peers or working. While extremely useful, smartphone apps also have drawbacks, as they can affect the security and privacy of users. Android devices hold a lot of personal data from users, including their social circles (e.g., contacts), usage patterns (e.g., app usage and visited websites) and their physical location. Like in most software products, Android apps often include third-party code (Software Development Kits or SDKs) to include functionality in the app without the need to develop it in-house. Android apps and third-party components embedded in them are often interested in accessing such data, as the online ecosystem is dominated by data-driven business models and revenue streams like advertising.
The research community has developed many methods and techniques for analyzing the privacy and security risks of mobile apps, mostly relying on two techniques: static code analysis and dynamic runtime analysis. Static analysis analyzes the code and other resources of an app to detect potential app behaviors. While this makes static analysis easier to scale, it has other drawbacks such as missing app behaviors when developers obfuscate the app’s code to avoid scrutiny. Furthermore, since static analysis only shows potential app behavior, this needs to be confirmed as it can also report false positives due to dead or legacy code. Dynamic analysis analyzes the apps at runtime to provide actual evidence of their behavior. However, these techniques are harder to scale as they need to be run on an instrumented device to collect runtime data. Similarly, there is a need to stimulate the app, simulating real inputs to examine as many code-paths as possible. While there are some automatic techniques to generate synthetic inputs, they have been shown to be insufficient.
In this thesis, we explore the benefits of combining static and dynamic analysis techniques to complement each other and reduce their limitations. While most previous work has often relied on using these techniques in isolation, we combine their strengths in different and novel ways that allow us to further study different privacy issues on the Android ecosystem. Namely, we demonstrate the potential of combining these complementary methods to study three inter-related issues:
A regulatory analysis of parental control apps. We use a novel methodology that relies on easy-to-scale static analysis techniques to pin-point potential privacy issues and violations of current legislation by Android apps and their embedded SDKs. We rely on the results from our static analysis to inform the way in which we manually exercise the apps, maximizing our ability to obtain real evidence of these misbehaviors. We study 46 publicly available apps and find instances of data collection and sharing without consent and insecure network transmissions containing personal data. We also see that these apps fail to properly disclose these practices in their privacy policy.
A security analysis of the unauthorized access to permission-protected data without user consent. We use a novel technique that combines the strengths of static and dynamic analysis, by first comparing the data sent by applications at runtime with the permissions granted to each app in order to find instances of potential unauthorized access to permission protected data. Once we have discovered the apps that are accessing personal data without permission, we statically analyze their code in order to discover covert- and side-channels used by apps and SDKs to circumvent the permission system. This methodology allows us to discover apps using the MAC address as a surrogate for location data, two SDKs using the external storage as a covert-channel to share unique identifiers and an app using picture metadata to gain unauthorized access to location data.
A novel SDK detection methodology that relies on obtaining signals observed both in the app’s code and static resources and during its runtime behavior. Then, we rely on a tree structure together with a confidence based system to accurately detect SDK presence without the need of any a priory knowledge and with the ability to discern whether a given SDK is part of legacy or dead code. We prove that this novel methodology can discover third-party SDKs with more accuracy than state-of-the-art tools both on a set of purpose-built ground-truth apps and on a dataset of 5k publicly available apps. With these three case studies, we are able to highlight the benefits of combining static and dynamic analysis techniques for the study of the privacy and security guarantees and risks of Android apps and third-party SDKs. The use of these techniques in isolation would not have allowed us to deeply investigate these privacy issues, as we would lack the ability to provide real evidence of potential breaches of legislation, to pin-point the specific way in which apps are leveraging cover and side channels to break Android’s permission system or we would be unable to adapt to an ever-changing ecosystem of Android third-party companies.Telematics EngineeringUniversidad Carlos III de Madrid, Spai