Search for: All records

Despite remarkable improvements, automatic speech recognition is susceptible to adversarial perturbations. Compared to standard machine learning architectures, these attacks are significantly more challenging, especially since the inputs to a speech recognition system are time series that contain both acoustic and linguistic properties of speech. Extracting all recognition-relevant information requires more complex pipelines and an ensemble of specialized components. Consequently, an attacker needs to consider the entire pipeline. In this paper, we present VENOMAVE, the first training- time poisoning attack against speech recognition. Similar to the predominantly studied evasion attacks, we pursue the same goal: leading the system to an incorrect and attacker-chosen transcription of a target audio waveform. In contrast to evasion attacks, however, we assume that the attacker can only manipulate a small part of the training data without altering the target audio waveform at runtime. We evaluate our attack on two datasets: TIDIGITS and Speech Commands. When poisoning less than 0.17% of the dataset, VENOMAVE achieves attack success rates of more than 80.0%, without access to the victim’s network architecture or hyperparameters. In a more realistic scenario, when the target audio waveform is played over the air in different rooms, VENOMAVE maintains a success rate of up to 73.3%. Finally, VENOMAVE achieves an attack transferability rate of 36.4% between two different model architectures. 

Security monitoring systems detect potentially malicious activities in IT infrastructures, by either looking for known signatures or for anomalous behaviors. Security operators investigate these events to determine whether they pose a threat to their organization. In many cases, a single event may be insufficient to determine whether certain activity is indeed malicious. Therefore, a security operator frequently needs to correlate multiple events to identify if they pose a real threat. Unfortunately, the vast number of events that need to be correlated often overload security operators, forcing them to ignore some events and, thereby, potentially miss attacks. This work studies how to automatically correlate security events and, thus, automate parts of the security operator workload. We design and evaluate DEEPCASE, a system that leverages the context around events to determine which events require further inspection. This approach reduces the number of events that need to be inspected. In addition, the context provides valuable insights into why certain events are classified as malicious. We show that our approach automatically filters 86.72% of the events and reduces the manual workload of security operators by 90.53%, while underestimating the risk of potential threats in less than 0.001% of cases. 

Malware analysis is an essential task to understand infection campaigns, the behavior of malicious codes, and possible ways to mitigate threats. Malware analysis also allows better assessment of attacker’s capabilities, techniques, and processes. Although a substantial amount of previous work provided a comprehensive analysis of the international malware ecosystem, research on regionalized, country, and population-specific malware campaigns have been scarce. Moving towards addressing this gap, we conducted a longitudinal (2012-2020) and comprehensive (encompassing an entire population of online banking users) study of MS Windows desktop malware that actually infected Brazilian bank’s users. We found that the Brazilian financial desktop malware has been evolving quickly: it started to make use of a variety of file formats instead of typical PE binaries, relied on native system resources, and abused obfuscation technique to bypass detection mechanisms. Our study on the threats targeting a significant population on the ecosystem of the largest and most populous country in Latin America can provide invaluable insights that may be applied to other countries’ user populations, especially those in the developing world that might face cultural peculiarities similar to Brazil’s. With this evaluation, we expect to motivate the security community/industry to seriously considering a deeper level of customization during the development of next generation anti-malware solutions, as well as to raise awareness towards regionalized and targeted Internet threats. 

Malware detection plays a vital role in computer security. Modern machine learning approaches have been centered around domain knowledge for extracting malicious features. However, many potential features can be used, and it is time consuming and difficult to manually identify the best features, especially given the diverse nature of malware. In this paper, we propose Neurlux, a neural network for malware detection. Neurlux does not rely on any feature engineering, rather it learns automatically from dynamic analysis reports that detail behavioral information. Our model borrows ideas from the field of document classification, using word sequences present in the reports to predict if a report is from a malicious binary or not. We investigate the learned features of our model and show which components of the reports it tends to give the highest importance. Then, we evaluate our approach on two different datasets and report formats, showing that Neurlux improves on the state of the art and can effectively learn from the dynamic analysis reports. Furthermore, we show that our approach is portable to other malware analysis environments and generalizes to different datasets. 

Machine learning techniques are widely used in addition to signatures and heuristics to increase the detection rate of anti-malware software, as they automate the creation of detection models, making it possible to handle an ever-increasing number of new malware samples. In order to foil the analysis of anti-malware systems and evade detection, malware uses packing and other forms of obfuscation. However, few realize that benign applications use packing and obfuscation as well, to protect intellectual property and prevent license abuse. In this paper, we study how machine learning based on static analysis features operates on packed samples. Malware researchers have often assumed that packing would prevent machine learning techniques from building effective classifiers. However, both industry and academia have published results that show that machine-learning-based classifiers can achieve good detection rates, leading many experts to think that classifiers are simply detecting the fact that a sample is packed, as packing is more prevalent in malicious samples. We show that, different from what is commonly assumed, packers do preserve some information when packing programs that is “useful” for malware classification. However, this information does not necessarily capture the sample’s behavior. We demonstrate that the signals extracted from packed executables are not rich enough for machine-learning-based models to (1) generalize their knowledge to operate on unseen packers, and (2) be robust against adversarial examples. We also show that a na¨ıve application of machine learning techniques results in a substantial number of false positives, which, in turn, might have resulted 

Binary static analysis has seen a recent surge in interest, due to a rise in analysis targets for which no other method is appropriate, such as, embedded firmware. This has led to the proposal of a number of binary static analysis tools and techniques, handling various kinds of programs, and answering different research questions. While static analysis tools that focus on binaries inherit the undecidability of static analysis, they bring with them other challenges, particularly in dealing with the aliasing of code and data pointers. These tools may tackle these challenges in different ways, but unfortunately, there is currently no concrete means of comparing their effectiveness at solving these central, problem-independent aspects of static analysis. In this paper, we propose a new method for creating a dataset of real-world programs, paired with the ground truth for static analysis. Our approach involves the injection of synthetic “facts” into a set of open-source programs, consisting of new variables and their possible values. The analyses’ goal is then to evaluate the possible values of these facts at certain program points. As the facts are injected randomly within an arbitrarily-large set of programs, the kinds of data flows that can be measured are widely-varied in size and complexity. We implemented this idea as a prototype system, AUTOFACTS, and used it to create a ground truth dataset of 29 programs, with various types and number of facts, resulting in a total of 2,088 binaries (with 72 versions for each program). To our knowledge, this is the first dataset aimed at the problem-independent evaluation of static analysis tools, and we contribute all code and the dataset itself to the community as open-source.