System call interposition is a widely used technique to trace and modify application behavior. Over the years, numerous interposition mechanisms have been proposed, each with distinct strengths and trade-offs. Recently, advances in binary rewriting—specifically targeting x86–64 syscall and sysenter instructions—have led to new techniques that take important steps forward, with some claiming to support general-purpose use.

Membership Inference Attacks (MIAs) can be conducted based on specific settings/assumptions and experience different limitations. In this paper, first, we provide a systematization of knowledge for all representative MIAs found in the literature. Second, we empirically evaluate and compare the MIA success rates achieved on Machine Learning (ML) models trained with some of the most common generalization techniques. Third, we examine the contribution of potential data leaks to successful MIAs. Fourth, we examine if the depth of Artificial Neural Networks (ANNs) affects MIA success rate and to what extent. For the experimental analysis, we focus solely on well-generalizable target models (various architectures trained on multiple datasets), having only black-box access to them. Our results suggest the following: (a) MIAs on well-generalizable targets suffer from significant limitations which undermine their practicality, (b) common generalization techniques result in ML models which are comparably robust against MIAs, (c) data leaks, although effective for overfitted models, do not facilitate MIAs in case of well-generalizable targets, (d) deep ANN architectures are not more vulnerable to MIAs compared to shallower ones or the opposite, and (e) well-generalizable models can be robust against MIAs even when not achieving state-of-the-art performance.

Model Inversion (MI) attacks, that aim to recover semantically meaningful reconstructions for each target class, have been extensively studied and demonstrated to be successful in the white-box setting. On the other hand, black-box MI attacks demonstrate low performance in terms of both effectiveness, i.e., reconstructing samples which are identifiable as their ground-truth, and efficiency, i.e., time or queries required for completing the attack process. Whether or not effective and efficient black-box MI attacks can be conducted on complex targets, such as Convolutional Neural Networks (CNNs), currently remains unclear. In this paper, we present a feasibility study in regards to the effectiveness and efficiency of MI attacks in the black-box setting. In this context, we introduce Deep-BMI (Deep Black-box Model Inversion), a framework that supports various black-box optimizers for conducting MI attacks on deep CNNs used for image recognition. Deep-BMI’s most efficient optimizer is based on an adaptive hill climbing algorithm, whereas its most effective optimizer is based on an evolutionary algorithm capable of performing an all-class attack and returning a diversity of images in a single run. For assessing the severity of this threat, we utilize all three evaluation approaches found in the literature. In particular, we (a) conduct a user study with human participants, (b) demonstrate our actual reconstructions along with their ground-truth, and (c) use relevant quantitative metrics. Surprisingly, our results suggest that black-box MI attacks, and for complex models, are comparable, in some cases, to those reported so far in the white-box setting.

Honeywords are false passwords associated with each user account. Using a honeyword to login sets off an alarm as a data breach has been detected. Existing approaches for detecting data breaches using honeywords suffer from the need of a trusted component to tell honey-words from the valid password. Once this trusted component is compromised, then honeywords can offer no assistance for mitigating or detecting a data breach. In this paper, we present Lethe, a honeywords-based data-breach detection system that requires no trusted components, other than a trusted bootstrap, and keeps limited transient state for verifying login attempts. Lethe is based on two fundamental principles. First, Lethe generates honeywords using a Machine Learning (ML) model, which constantly evolves. This means that an attacker that compromises the Honeyword Generation Technique (HGT) cannot reproduce the same set of honeywords, and thus cannot tell which password was used as the initial generator. In particular, Lethe is the first system that allows an attacker to fully compromise the HGT without affecting the security of already generated honeywords. Second, Lethe is not aware of the valid password. In fact, for Lethe the only one that knows the actual password is the user that selected it in the first place. Lethe records login events, but without storing anywhere the password used. These login events can be further replayed in another server, which can check if, for a particular user, there were at least two different passwords used and therefore detect a data breach. Lethe allows the detection of a data breach deterministically and not probabilistically as similar approaches do. Additionally, Lethe allows detecting data breaches that are associated with rarely used accounts. Lethe can signal an alarm even if a user account that has logged in just once with the system is compromised. This is in contrast to other efforts that require legitimate users to authenticate with the system, after the attacker has done so, for detecting the breach. To demonstrate the effectiveness of Lethe, we provide a fully functional prototype, along with the ML-based HGT, and assess the provided security with a set of diverse attackers.

Honeywords are false passwords injected in a database for detecting password leakage. Generating honeywords is a challenging problem due to the various assumptions about the adversary's knowledge as well as users' password-selection behaviour. The success of a Honeywords Generation Technique (HGT) lies on the resulting honeywords; the method fails if an adversary can easily distinguish the real password. In this paper, we propose HoneyGen, a practical and highly robust HGT that produces realistic looking honeywords. We do this by leveraging representation learning techniques to learn useful and explanatory representations from a massive collection of unstructured data, i.e., each operator's password database. We perform both a quantitative and qualitative evaluation of our framework using the state-of-the-art metrics. Our results suggest that HoneyGen generates high-quality honeywords that cause sophisticated attackers to achieve low distinguishing success rates.

Internet services heavily rely on CAPTCHAs for determining whether or not a user is a human being. The recent advances in ML and AI make the efficacy of CAPTCHAs in strengthening Internet services against bots questionable. In this paper, we conduct a systematic analysis and classification of the state-of-the-art ML-based techniques for the automated text-based CAPTCHA breaking problem. The current state and robustness of text-based CAPTCHAs as are utilized by modern Internet applications, against ML-based automated breaking tools, is examined and reported. Our study suggests that ML can be very effective in increasing: (a) accuracy, (b) speed, and (c) abstraction in CAPTCHA solving. Especially, as far as (c) is concerned, ML-based techniques are easier to be applied in different classes of text-based CAPTCHA schemes. To assess the importance of ML in breaking CAPTCHAs, we build our own ML-only classifiers. Surprisingly, an ML-only approach for solving CAPTCHAs is not sufficient. Overall, our study suggests that fundamentally different ways of conducting reverse Turing test, that will be painless for legitimate users (i.e., humans) but at the same time challenging for automated systems (i.e., software), should be considered for ensuring the healthy operation of current Internet services.