Malware classification is a specific and refined task within the broader malware detection problem. Effective classification aids in understanding attack techniques and developing robust defenses, ensuring application security and timely mitigation of software vulnerabilities. The dynamic nature of malware demands adaptive classification techniques that can handle the continuous emergence of new families. Traditionally, this is done by retraining models on all historical samples, which requires significant resources in terms of time and storage. An alternative approach is Class-Incremental Learning (CIL), which focuses on progressively learning new classes (malware families) while preserving knowledge from previous training steps. However, CIL assumes that each class appears only once in training and is not revisited, an assumption that does not hold for malware families, which often persist across multiple time intervals. This leads to shifts in the data distribution for the same family over time, a challenge that is not addressed by traditional CIL methods. We formulate this problem as Temporal-Incremental Malware Learning (TIML), which adapts to these shifts and effectively classifies new variants. To support this, we organize the MalNet dataset, consisting of over a million entries of Android malware data collected over a decade, in chronological order. We first adapt state-of-the-art CIL approaches to meet TIML’s requirements, serving as baseline methods. Then, we propose a novel multimodal TIML approach that leverages multiple malware modalities for improved performance. Extensive evaluations show that our TIML approaches outperform traditional CIL methods and demonstrate the feasibility of periodically updating malware classifiers at a low cost. This process is efficient and requires minimal storage and computational resources, with only a slight dip in performance compared to full retraining with historical data.

Recent advancements in ML and DL have significantly improved Android malware detection, yet many methodologies still rely on basic static analysis, bytecode, or function call graphs that often fail to capture complex malicious behaviors. DexBERT, a pre-trained BERT-like model tailored for Android representation learning, enriches class-level representations by analyzing Smali code extracted from APKs. However, its functionality is constrained by its inability to process multiple Smali classes simultaneously. This paper introduces DetectBERT, which integrates correlated Multiple Instance Learning (c-MIL) with DexBERT to handle the high dimensionality and variability of Android malware, enabling effective app-level detection. By treating class-level features as instances within MIL bags, DetectBERT aggregates these into a comprehensive app-level representation. Our evaluation demonstrates that DetectBERT not only surpasses existing state-of-the-art detection methods but also adapts to evolving malware threats. Moreover, the versatility of the DetectBERT framework holds promising potential for broader applications in app-level analysis and other software engineering tasks, offering new avenues for research and development.

In 2016, we released AndroZoo, a continuously expanding dataset of Android applications that aggregates apps from various sources, including the official Google Play app market. As of today, Andro-Zoo contains approximately 24 million APK files, making it, to the best of our knowledge, the most extensive dataset of Android APKs accessible to the research community. Currently, an average of 500 000 APKs are downloaded every day, with our initial MSR paper counting more than 880 citations on Google Scholar.

The popularity of Android OS has made it an appealing target for malware developers. To evade detection, including by ML-based techniques, attackers invest in creating malware that closely resemble legitimate apps, challenging the state of the art with difficult-to-detect samples. In this paper, we propose Guided Retraining, a supervised representation learning-based method for boosting the performance of malware detectors. To that end, we first split the experimental dataset into subsets of “easy” and “difficult” samples, where difficulty is associated to the prediction probabilities yielded by a malware detector. For the subset of “easy” samples, the base malware detector is used to make the final predictions since the error rate on that subset is low by construction. Our work targets the second subset containing “difficult” samples, for which the probabilities are such that the classifier is not confident on the predictions, which have high error rates. We apply our Guided Retraining method on these difficult samples to improve their classification. Guided Retraining leverages the correct predictions and the errors made by the base malware detector to guide the retraining process. Guided Retraining learns new embeddings of the difficult samples using Supervised Contrastive Learning and trains an auxiliary classifier for the final predictions. We validate our method on four state-of-the-art Android malware detection approaches using over 265k malware and benign apps. Experimental results show that Guided Retraining can boost state-of-the-art detectors by eliminating up to 45.19% of the prediction errors that they make on difficult samples. We note furthermore that our method is generic and designed to enhance the performance of binary classifiers for other tasks beyond Android malware detection.

Abstract Research on Android malware detection based on Machine learning has been prolific in recent years. In this paper, we show, through a large-scale evaluation of four state-of-the-art approaches that their achieved performance fluctuates when applied to different datasets. Combining existing approaches appears as an appealing method to stabilise performance. We therefore proceed to empirically investigate the effect of such combinations on the overall detection performance. In our study, we evaluated 22 methods to combine feature sets or predictions from the state-of-the-art approaches. Our results showed that no method has significantly enhanced the detection performance reported by the state-of-the-art malware detectors. Nevertheless, the performance achieved is on par with the best individual classifiers for all settings. Overall, we conduct extensive experiments on the opportunity to combine state-of-the-art detectors. Our main conclusion is that combining state-of-the-art malware detectors leads to a stabilisation of the detection performance, and a research agenda on how they should be combined effectively is required to boost malware detection. All artefacts of our large-scale study (i.e., the dataset of $\sim $ ∼ 0.5 million apks and all extracted features) are made available for replicability.

The automation of an increasingly large number of software engineering tasks is becoming possible thanks to Machine Learning (ML). One foundational building block in the application of ML to software artifacts is therepresentationof these artifacts (e.g., source code or executable code) into a form that is suitable for learning. Traditionally, researchers and practitioners have relied on manually selected features, based on expert knowledge, for the task at hand. Such knowledge is sometimes imprecise and generally incomplete. To overcome this limitation, many studies have leveraged representation learning, delegating to ML itself the job of automatically devising suitable representations and selections of the most relevant features. Yet, in the context of Android problems, existing models are either limited to coarse-grained whole-app level (e.g., apk2vec) or conducted for one specific downstream task (e.g., smali2vec). Thus, the produced representation may turn out to be unsuitable for fine-grained tasks or cannot generalize beyond the task that they have been trained on. Our work is part of a new line of research that investigates effective, task-agnostic, and fine-grained universal representations of bytecode to mitigate both of these two limitations. Such representations aim to capture information relevant to various low-level downstream tasks (e.g., at the class-level). We are inspired by the field of Natural Language Processing, where the problem of universal representation was addressed by building Universal Language Models, such as BERT, whose goal is to capture abstract semantic information about sentences, in a way that is reusable for a variety of tasks. We propose DexBERT, a BERT-like Language Model dedicated to representing chunks of DEX bytecode, the main binary format used in Android applications. We empirically assess whether DexBERT is able to model the DEXlanguageand evaluate the suitability of our model in three distinct class-level software engineering tasks: Malicious Code Localization, Defect Prediction, and Component Type Classification. We also experiment with strategies to deal with the problem of catering to apps having vastly different sizes, and we demonstrate one example of using our technique to investigate what information is relevant to a given task.

peer reviewed

Among the reported solutions to the class imbalance issue, the undersampling approaches, which remove instances of insignificant samples from the majority class, are quite prevalent. However, the undersampling approaches may discard significant patterns in the datasets. A prototype, which is always an actual sample from the data, represents a group of samples in the dataset. Our hypothesis is that prototypes can fill the missing significant patterns that are discarded by undersampling methods and help to improve model performance. To confirm our intuition, we articulate prototypes to undersampling methods in the machine learning pipeline. We show that there is a statistically significant difference between the AUPR and AUROC results of undersampling methods and our approach.

Native code is now commonplace within Android app packages where it co-exists and interacts with Dex bytecode through the Java Native Interface to deliver rich app functionalities. Yet, state-of-the-art static analysis approaches have mostly overlooked the presence of such native code, which, however, may implement some key sensitive, or even malicious, parts of the app behavior. This limitation of the state of the art is a severe threat to validity in a large range of static analyses that do not have a complete view of the executable code in apps. To address this issue, we propose a new advance in the ambitious research direction of building a unified model of all code in Android apps. The JuCify approach presented in this paper is a significant step towards such a model, where we extract and merge call graphs of native code and bytecode to make the final model readily-usable by a common Android analysis framework: in our implementation, JuCify builds on the Soot internal intermediate representation. We performed empirical investigations to highlight how, without the unified model, a significant amount of Java methods called from the native code are "unreachable" in apps' call-graphs, both in goodware and malware. Using JuCify, we were able to enable static analyzers to reveal cases where malware relied on native code to hide invocation of payment library code or of other sensitive code in the Android framework. Additionally, JuCify's model enables state-of-the-art tools to achieve better precision and recall in detecting data leaks through native code. Finally, we show that by using JuCify we can find sensitive data leaks that pass through native code.

Pre-trained Language Models such as BERT have become ubiquitous in NLP where they have achieved state-of-the-art performance in most NLP tasks. While these models are readily available for English and other widely spoken languages, they remain scarce for low-resource languages such as Luxembourgish. In this paper, we present LuxemBERT, a BERT model for the Luxembourgish language that we create using the following approach: we augment the pre-training dataset by considering text data from a closely related language that we partially translate using a simple and straightforward method. We are then able to produce the LuxemBERT model, which we show to be effective for various NLP tasks: it outperforms a simple baseline built with the available Luxembourgish text data as well the multilingual mBERT model, which is currently the only option for transformer-based language models in Luxembourgish. Furthermore, we present datasets for various downstream NLP tasks that we created for this study and will make available to researchers on request.

This paper investigates the various technical and non-technical tools and techniques that software developers use to build and disseminate crypto mining apps on Android devices. Our study of 346 potential Android mining apps, collected between April 2019 and May 2022, has revealed the presence of more than ten mining apps on the Google Play Store, with at least half of those still available at the time of writing this (June 2022). We observed that many of those mining apps do not conceal their usage of the device's resource for mining which is considered a violation of the store's policies for developers. We estimate that more than ten thousand users have run mining apps downloaded directly from the Google Play Store, which puts the supposedly “stringent” vetting process into question. Furthermore, we prove that covert mining apps tend to be embedded into supposedly free versions of premium apps or pose as utility apps that provide valuable features to users. Finally, we empirically demonstrate that cryptojacking apps' resource consumption and malicious behavior could be insignificant. We presume that typical users, even though they might be running a mobile antivirus solution, could execute a mining app for an extended period without being alerted. We expect our results to inform the various actors involved in the security of Android devices against the lingering threat of cryptojacking and help them better assess the problem.

Machine learning advances have been extensively explored for implementing large-scale malware detection. When reported in the literature, performance evaluation of machine learning based detectors generally focuses on highlighting the ratio of samples that are correctly or incorrectly classified, overlooking essential questions on why/how the learned models can be demonstrated as reliable. In the Android ecosystem, several recent studies have highlighted how evaluation setups can carry biases related to datasets or evaluation methodologies. Nevertheless, there is little work attempting to dissect the produced model to provide some understanding of its intrinsic characteristics. In this work, we fill this gap by performing a comprehensive analysis of a state-of-the-art Android malware detector, namely DREBIN, which constitutes today a key reference in the literature. Our study mainly targets an in-depth understanding of the classifier characteristics in terms of (1) which features actually matter among the hundreds of thousands that DREBIN extracts, (2) whether the high scores of the classifier are dependent on the dataset age, and (3) whether DREBIN’s explanations are consistent within malware families, among others. Overall, our tentative analysis provides insights into the discriminatory power of the feature set used by DREBIN to detect malware. We expect our findings to bring about a systematisation of knowledge for the community.

Abstract A well-known curse of computer security research is that it often produces systems that, while technically sound, fail operationally. To overcome this curse, the community generally seeks to assess proposed systems under a variety of settings in order to make explicit every potential bias. In this respect, recently, research achievements on machine learning based malware detection are being considered for thorough evaluation by the community. Such an effort of comprehensive evaluation supposes first and foremost the possibility to perform an independent reproduction study in order to sharpen evaluations presented by approaches’ authors. The question Can published approaches actually be reproduced? thus becomes paramount despite the little interest such mundane and practical aspects seem to attract in the malware detection field. In this paper, we attempt a complete reproduction of five Android Malware Detectors from the literature and discuss to what extent they are “reproducible”. Notably, we provide insights on the implications around the guesswork that may be required to finalise a working implementation. Finally, we discuss how barriers to reproduction could be lifted, and how the malware detection field would benefit from stronger reproducibility standards—like many various fields already have.

Abstract Detecting vulnerabilities in software is a constant race between development teams and potential attackers. While many static and dynamic approaches have focused on regularly analyzing the software in its entirety, a recent research direction has focused on the analysis of changes that are applied to the code. VCCFinder is a seminal approach in the literature that builds on machine learning to automatically detect whether an incoming commit will introduce some vulnerabilities. Given the influence of VCCFinder in the literature, we undertake an investigation into its performance as a state-of-the-art system. To that end, we propose to attempt a replication study on the VCCFinder supervised learning approach. The insights of our failure to replicate the results reported in the original publication informed the design of a new approach to identify vulnerability-contributing commits based on a semi-supervised learning technique with an alternate feature set. We provide all artefacts and a clear description of this approach as a new reproducible baseline for advancing research on machine learning-based identification of vulnerability-introducing commits.

Due to the convenience of access-on-demand to information and business solutions, mobile apps have become an important asset in the digital world. In the context of the COVID-19 pandemic, app developers have joined the response effort in various ways by releasing apps that target different user bases (e.g., all citizens or journalists), offer different services (e.g., location tracking or diagnostic-aid), provide generic or specialized information, etc. While many apps have raised some concerns by spreading misinformation or even malware, the literature does not yet provide a clear landscape of the different apps that were developed. In this study, we focus on the Android ecosystem and investigate Covid-related Android apps. In a best-effort scenario, we attempt to systematically identify all relevant apps and study their characteristics with the objective to provide a first taxonomy of Covid-related apps, broadening the relevance beyond the implementation of contact tracing. Overall, our study yields a number of empirical insights that contribute to enlarge the knowledge on Covid-related apps: (1) Developer communities contributed rapidly to the COVID-19, with dedicated apps released as early as January 2020; (2) Covid-related apps deliver digital tools to users (e.g., health diaries), serve to broadcast information to users (e.g., spread statistics), and collect data from users (e.g., for tracing); (3) Covid-related apps are less complex than standard apps; (4) they generally do not seem to leak sensitive data; (5) in the majority of cases, Covid-related apps are released by entities with past experience on the market, mostly official government entities or public health organizations.

Named Entity Recognition (NER) is a fundamental Natural Language Processing (NLP) task and has remained an active research field.In recent years, transformer models and more specifically the BERT model developed at Google revolutionised the field of NLP.While the performance of transformer-based approaches such as BERT has been studied for NER, there has not yet been a study for the fine-grained Named Entity Recognition (FG-NER) task.In this paper, we compare three transformer-based models (BERT, RoBERTa, and XLNet) to two non-transformer-based models (CRF and BiLSTM-CNN-CRF).Furthermore, we apply each model to a multitude of distinct domains.We find that transformer-based models incrementally outperform the studied non-transformer-based models in most domains with respect to the F1 score.Furthermore, we find that the choice of domain significantly influenced the performance regardless of the respective data size or the model chosen.

We present a growing collection of Android Applications collected from several sources, including the official Google Play app market. Our dataset, AndroZoo, currently contains more than three million apps, each of which has been analysed by tens of different Antivirus products to know which applications are detected as Malware. We provide this dataset to contribute to ongoing research efforts, as well as to enable new potential research topics on Android Apps. By releasing our dataset to the research community, we also aim at encouraging our fellow researchers to engage in reproducible experiments.

We discuss the capability of a new feature set for malware detection based on potential component leaks (PCLs). PCLs are defined as sensitive data-flows that involve Android inter-component communications. We show that PCLs are common in Android apps and that malicious applications indeed manipulate significantly more PCLs than benign apps. Then, we evaluate a machine learning-based approach relying on PCLs. Experimental validations show high performance for identifying malware, demonstrating that PCLs can be used for discriminating malicious apps from benign apps.

To address the issue of malware detection, researchers have recently started to investigate the capabilities of machine-learning techniques for proposing effective approaches. Several promising results were recorded in the literature, many approaches being assessed with the common "10-Fold cross validation" scheme. This paper revisits the purpose of malware detection to discuss the adequacy of the "10-Fold" scheme for validating techniques that may not perform well in reality. To this end, we have devised several Machine Learning classifiers that rely on a novel set of features built from applications' CFGs. We use a sizeable dataset of over 50,000 Android applications collected from sources where state-of-the art approaches have selected their data. We show that our approach outperforms existing machine learning-based approaches. However, this high performance on usual-size datasets does not translate in high performance in the wild.

We consider in this paper the analysis of a large set of malware and benign applications from the Android ecosystem. Although a large body of research work has dealt with Android malware over the last years, none has addressed it from a forensic point of view. After collecting over 500,000 applications from user markets and research repositories, we perform an analysis that yields precious insights on the writing process of Android malware. This study also explores some strange artifacts in the datasets, and the divergent capabilities of state-of-the-art antivirus to recognize/define malware. We further highlight some major weak usage and misunderstanding of Android security by the criminal community and show some patterns in their operational flow. Finally, using insights from this analysis, we build a naive malware detection scheme that could complement existing anti virus software.

Recently, the Android platform has seen its number of malicious applications increased sharply. Motivated by the easy application submission process and the number of alternative market places for distributing Android applications, rogue authors are developing constantly new malicious programs. While current anti-virus software mainly relies on signature detection, the issue of alternative malware detection has to be addressed. In this paper, we present a feature based detection mechanism relying on opcode-sequences combined with machine learning techniques. We assess our tool on both a reference dataset known as Genome Project as well as on a wider sample of 40,000 applications retrieved from the Google Play Store.