arXiv subjects

Sjouke Mauw

Publications and source records attributed to Sjouke Mauw.

At least 19 recordsLinked to original sources

arXiv2026

How does the optimizer implicitly bias the model merging loss landscape?

Model merging combines independent solutions with different capabilities into a single one while maintaining the same inference cost. Two popular approaches are linear interpolation, which simply averages multiple model weights, and task arithmetic, which combines task vectors obtained by the difference between finetuned and base models. While useful in practice, what properties make merging effective are poorly understood. This paper explores how the optimization dynamics affect the loss landscape geometry and its impact on merging success. We show that a single quantity -- the effective noise scale -- unifies the impact of different optimizer components on model merging. Across architectures and datasets, merging success is a non-monotonic function of the effective noise scale, with a distinct optimum. Decomposing this quantity, we find that larger learning rates, stronger weight decay, smaller batch sizes, and data augmentation all independently modulate the effective noise scale and exhibit the same qualitative trend. Unlike prior work connecting optimizer noise to the flatness or generalization of individual minima, we show that it also affects the global loss landscape, predicting when independently trained solutions can be successfully merged. Our findings broaden the understanding of how optimization shapes the loss landscape geometry and its consequences for model merging, suggesting that training dynamics could be further manipulated to improve model merging.

cs.LG↗

arXiv2026

Unlinkability and History Preserving Bisimilarity

An ever-increasing number of critical infrastructures rely heavily on the assumption that security protocols satisfy a wealth of requirements. Hence, the importance of certifying e.g., privacy properties using methods that are better at detecting attacks can hardly be overstated. This paper scrutinises the "unlinkability" privacy property using relations equating behaviours that cannot be distinguished by attackers. Starting from the observation that some reasonable design choice can lead to formalisms missing attacks, we draw attention to a classical concurrent semantics accounting for relationship between past events, and show that there are concurrency-aware semantics that can discover attacks on all protocols we consider.More precisely, we focus on protocols where trace equivalence is known to miss attacks that are observable using branching-time equivalences. We consider the impact of three dimensions: design decisions made by the programmer specifying an unlinkability problem (style), semantics respecting choices during execution (branching-time), and semantics sensitive to concurrency (non-interleaving), and discover that reasonable styles miss attacks unless we give attackers enough power to observe choices and concurrency. Our main contribution is to draw attention to how a popular concurrent semantics -- history-preserving bisimilarity -- when defined for the non-interleaving applied $π$-calculus, can discover attacks on all protocols we consider, regardless of the choice of style. Furthermore, we can describe all such attacks using a novel modal logic that is hence suitable to formally certify attacks on privacy properties.

cs.CR↗

arXiv2025

Bits for Privacy: Evaluating Post-Training Quantization via Membership Inference

Deep neural networks are widely deployed with quantization techniques to reduce memory and computational costs by lowering the numerical precision of their parameters. While quantization alters model parameters and their outputs, existing privacy analyses primarily focus on full-precision models, leaving a gap in understanding how bit-width reduction can affect privacy leakage. We present the first systematic study of the privacy-utility relationship in post-training quantization (PTQ), a versatile family of methods that can be applied to pretrained models without further training. Using membership inference attacks as our evaluation framework, we analyze three popular PTQ algorithms-AdaRound, BRECQ, and OBC-across multiple precision levels (4-bit, 2-bit, and 1.58-bit) on CIFAR-10, CIFAR-100, and TinyImageNet datasets. Our findings consistently show that low-precision PTQs can reduce privacy leakage. In particular, lower-precision models demonstrate up to an order of magnitude reduction in membership inference vulnerability compared to their full-precision counterparts, albeit at the cost of decreased utility. Additional ablation studies on the 1.58-bit quantization level show that quantizing only the last layer at higher precision enables fine-grained control over the privacy-utility trade-off. These results offer actionable insights for practitioners to balance efficiency, utility, and privacy protection in real-world deployments.

cs.LG↗

arXiv2025

Spurious Privacy Leakage in Neural Networks

Neural networks trained on real-world data often exhibit biases while simultaneously being vulnerable to privacy attacks aimed at extracting sensitive information. Despite extensive research on each problem individually, their intersection remains poorly understood. In this work, we investigate the privacy impact of spurious correlation bias. We introduce \emph{spurious privacy leakage}, a phenomenon in which spurious groups are significantly more vulnerable to privacy attacks than non-spurious groups. We observe that privacy disparity between groups increases in tasks with simpler objectives (e.g. fewer classes) due to spurious features. Counterintuitively, we demonstrate that spurious robust methods, designed to reduce spurious bias, fail to mitigate privacy disparity. Our analysis reveals that this occurs because robust methods can reduce reliance on spurious features for prediction, but do not prevent their memorization during training. Finally, we systematically compare the privacy of different model architectures trained with spurious data, demonstrating that, contrary to previous work, architectural choice can affect privacy evaluation.

cs.LG↗

arXiv2025

Empirical Evaluation of Memory-Erasure Protocols

Software-based memory-erasure protocols are two-party communication protocols where a verifier instructs a computational device to erase its memory and send a proof of erasure. They aim at guaranteeing that low-cost IoT devices are free of malware by putting them back into a safe state without requiring secure hardware or physical manipulation of the device. Several software-based memory-erasure protocols have been introduced and theoretically analysed. Yet, many of them have not been tested for their feasibility, performance and security on real devices, which hinders their industry adoption. This article reports on the first empirical analysis of software-based memory-erasure protocols with respect to their security, erasure guarantees, and performance. The experimental setup consists of 3 modern IoT devices with different computational capabilities, 7 protocols, 6 hash-function implementations, and various performance and security criteria. Our results indicate that existing software-based memory-erasure protocols are feasible, although slow devices may take several seconds to erase their memory and generate a proof of erasure. We found that no protocol dominates across all empirical settings, defined by the computational power and memory size of the device, the network speed, and the required level of security. Interestingly, network speed and hidden constants within the protocol specification played a more prominent role in the performance of these protocols than anticipated based on the related literature. We provide an evaluation framework that, given a desired level of security, determines which protocols offer the best trade-off between performance and erasure guarantees.

cs.CR↗

arXiv2024

Automated generation of attack trees with optimal shape and labelling

This article addresses the problem of automatically generating attack trees that soundly and clearly describe the ways the system can be attacked. Soundness means that the attacks displayed by the attack tree are indeed attacks in the system; clarity means that the tree is efficient in communicating the attack scenario. To pursue clarity, we introduce an attack-tree generation algorithm that minimises the tree size and the information length of its labels without sacrificing correctness. We achieve this by i) introducing a system model that allows to reason about attacks and goals in an efficient manner, and ii) by establishing a connection between the problem of factorising algebraic expressions and the problem of minimising the tree size. To the best of our knowledge, we introduce the first attack-tree generation framework that optimises the labelling and shape of the generated trees, while guaranteeing their soundness with respect to a system specification.

cs.CR↗

arXiv2024

Software-Based Memory Erasure with relaxed isolation requirements: Extended Version

A Proof of Secure Erasure (PoSE) is a communication protocol where a verifier seeks evidence that a prover has erased its memory within the time frame of the protocol execution. Designers of PoSE protocols have long been aware that, if a prover can outsource the computation of the memory erasure proof to another device, then their protocols are trivially defeated. As a result, most software-based PoSE protocols in the literature assume that provers are isolated during the protocol execution, that is, provers cannot receive help from a network adversary. Our main contribution is to show that this assumption is not necessary. We introduce formal models for PoSE protocols playing against provers aided by external conspirators and develop three PoSE protocols that we prove secure in this context. We reduce the requirement of isolation to the more realistic requirement that the communication with the external conspirator is relatively slow. Software-based protocols with such relaxed isolation assumptions are especially pertinent for low-end devices, where it is too costly to deploy sophisticated protection methods.

cs.CR↗

arXiv2023

Provably Unlinkable Smart Card-based Payments

The most prevalent smart card-based payment method, EMV, currently offers no privacy to its users. Transaction details and the card number are sent in cleartext, enabling the profiling and tracking of cardholders. Since public awareness of privacy issues is growing and legislation, such as GDPR, is emerging, we believe it is necessary to investigate the possibility of making payments anonymous and unlinkable without compromising essential security guarantees and functional properties of EMV. This paper draws attention to trade-offs between functional and privacy requirements in the design of such a protocol. We present the UTX protocol - an enhanced payment protocol satisfying such requirements, and we formally certify key security and privacy properties using techniques based on the applied pi-calculus.

cs.CR↗

arXiv2022

Unlinkability of an Improved Key Agreement Protocol for EMV 2nd Gen Payments

To address known privacy problems with the EMV standard, EMVCo have proposed a Blinded Diffie-Hellman key establishment protocol, which is intended to be part of a future 2nd Gen EMV protocol. We point out that active attackers were not previously accounted for in the privacy requirements of this proposal protocol, and demonstrate that an active attacker can compromise unlinkability within a distance of 100cm. Here, we adopt a strong definition of unlinkability that does account for active attackers and propose an enhancement of the protocol proposed by EMVCo. We prove that our protocol does satisfy strong unlinkability, while preserving authentication.

cs.CR↗

arXiv2022

Modelling Agent-Skipping Attacks in Message Forwarding Protocols

Message forwarding protocols are protocols in which a chain of agents handles transmission of a message. Each agent forwards the received message to the next agent in the chain. For example, TLS middleboxes act as intermediary agents in TLS, adding functionality such as filtering or compressing data. In such protocols, an attacker may attempt to bypass one or more intermediary agents. Such an agent-skipping attack can the violate security requirements of the protocol. Using the multiset rewriting model in the symbolic setting, we construct a comprehensive framework of such path protocols. In particular, we introduce a set of security goals related to path integrity: the notion that a message faithfully travels through participants in the order intended by the initiating agent. We perform a security analysis of several such protocols, highlighting key attacks on modern protocols.

cs.CR↗

arXiv2021

Discovering ePassport Vulnerabilities using Bisimilarity

We uncover privacy vulnerabilities in the ICAO 9303 standard implemented by ePassports worldwide. These vulnerabilities, confirmed by ICAO, enable an ePassport holder who recently passed through a checkpoint to be reidentified without opening their ePassport. This paper explains how bisimilarity was used to discover these vulnerabilities, which exploit the BAC protocol - the original ICAO 9303 standard ePassport authentication protocol - and remains valid for the PACE protocol, which improves on the security of BAC in the latest ICAO 9303 standards. In order to tackle such bisimilarity problems, we develop here a chain of methods for the applied $π$-calculus including a symbolic under-approximation of bisimilarity, called open bisimilarity, and a modal logic, called classical FM, for describing and certifying attacks. Evidence is provided to argue for a new scheme for specifying such unlinkability problems that more accurately reflects the capabilities of an attacker.

cs.CR↗

arXiv2020

Preventing active re-identification attacks on social graphs via sybil subgraph obfuscation

This paper addresses active re-identification attacks in the context of privacy-preserving social graph publication. Active attacks are those where the adversary can leverage fake accounts, a.k.a. sybil nodes, to enforce structural patterns that can be used to re-identify their victims on anonymised graphs. In this paper we present a new probabilistic interpretation of this type of attacks. Unlike previous privacy properties, which model the protection from active adversaries as the task of making victim nodes indistinguishable in terms of their fingerprints with respect to all potential attackers, our new formulation introduces a more complete view, where the attack is countered by jointly preventing the attacker from retrieving the set of sybil nodes, and from using these sybil nodes for re-identifying the victims. Under the new formulation, we show that the privacy property $k$-symmetry, introduced in the context of passive attacks, provides a sufficient condition for the protection against active re-identification attacks leveraging an arbitrary number of sybil nodes. Moreover, we show that the algorithm K-Match, originally devised for efficiently enforcing the related notion of $k$-automorphism, also guarantees $k$-symmetry. Empirical results on several collections of synthetic graphs corroborate that our approach allows, for the first time, to publish anonymised social graphs (with formal privacy guarantees) that effectively resist the strongest active re-identification attack reported in the literature, even when it leverages a large number of sybil nodes.

cs.SI↗

arXiv2020

ÆGIS: Shielding Vulnerable Smart Contracts Against Attacks

In recent years, smart contracts have suffered major exploits, costing millions of dollars. Unlike traditional programs, smart contracts are deployed on a blockchain. As such, they cannot be modified once deployed. Though various tools have been proposed to detect vulnerable smart contracts, the majority fails to protect vulnerable contracts that have already been deployed on the blockchain. Only very few solutions have been proposed so far to tackle the issue of post-deployment. However, these solutions suffer from low precision and are not generic enough to prevent any type of attack. In this work, we introduce ÆGIS, a dynamic analysis tool that protects smart contracts from being exploited during runtime. Its capability of detecting new vulnerabilities can easily be extended through so-called attack patterns. These patterns are written in a domain-specific language that is tailored to the execution model of Ethereum smart contracts. The language enables the description of malicious control and data flows. In addition, we propose a novel mechanism to streamline and speed up the process of managing attack patterns. Patterns are voted upon and stored via a smart contract, thus leveraging the benefits of tamper-resistance and transparency provided by the blockchain. We compare ÆGIS to current state-of-the-art tools and demonstrate that our solution achieves higher precision in detecting attacks. Finally, we perform a large-scale analysis on the first 4.5 million blocks of the Ethereum blockchain, thereby confirming the occurrences of well reported and yet unreported attacks in the wild.

cs.CR↗

arXiv2019

Active Re-identification Attacks on Periodically Released Dynamic Social Graphs

Active re-identification attacks pose a serious threat to privacy-preserving social graph publication. Active attackers create fake accounts to build structural patterns in social graphs which can be used to re-identify legitimate users on published anonymised graphs, even without additional background knowledge. So far, this type of attacks has only been studied in the scenario where the inherently dynamic social graph is published once. In this paper, we present the first active re-identification attack in the more realistic scenario where a dynamic social graph is periodically published. The new attack leverages tempo-structural patterns for strengthening the adversary. Through a comprehensive set of experiments on real-life and synthetic dynamic social graphs, we show that our new attack substantially outperforms the most effective static active attack in the literature by increasing the success probability of re-identification by more than two times and efficiency by almost 10 times. Moreover, unlike the static attack, our new attack is able to remain at the same level of effectiveness and efficiency as the publication process advances. We conduct a study on the factors that may thwart our new attack, which can help design graph anonymising methods with a better balance between privacy and utility.

cs.SI↗

arXiv2019

Publishing Community-Preserving Attributed Social Graphs with a Differential Privacy Guarantee

We present a novel method for publishing differentially private synthetic attributed graphs. Unlike preceding approaches, our method is able to preserve the community structure of the original graph without sacrificing the ability to capture global structural properties. Our proposal relies on C-AGM, a new community-preserving generative model for attributed graphs. We equip C-AGM with efficient methods for attributed graph sampling and parameter estimation. For the latter, we introduce differentially private computation methods, which allow us to release community-preserving synthetic attributed social graphs with a strong formal privacy guarantee. Through comprehensive experiments, we show that our new model outperforms its most relevant counterparts in synthesising differentially private attributed social graphs that preserve the community structure of the original graph, as well as degree sequences and clustering coefficients.

cs.SI↗

arXiv2019

Attribute Evaluation on Attack Trees with Incomplete Information

Attack trees are considered a useful tool for security modelling because they support qualitative as well as quantitative analysis. The quantitative approach is based on values associated to each node in the tree, expressing, for instance, the minimal cost or probability of an attack. Current quantitative methods for attack trees allow the analyst to, based on an initial assignment of values to the leaf nodes, derive the values of the higher nodes in the tree. In practice, however, it shows to be very difficult to obtain reliable values for all leaf nodes. The main reasons are that data is only available for some of the nodes, that data is available for intermediate nodes rather than for the leaf nodes, or even that the available data is inconsistent. We address these problems by developing a generalisation of the standard bottom-up calculation method in three ways. First, we allow initial attributions of non-leaf nodes. Second, we admit additional relations between attack steps beyond those provided by the underlying attack tree semantics. Third, we support the calculation of an approximative solution in case of inconsistencies. We illustrate our method, which is based on constraint programming, by a comprehensive case study.

cs.CR↗

arXiv2018

Robust active attacks on social graphs

In order to prevent the disclosure of privacy-sensitive data, such as names and relations between users, social network graphs have to be anonymised before publication. Naive anonymisation of social network graphs often consists in deleting all identifying information of the users, while maintaining the original graph structure. Various types of attacks on naively anonymised graphs have been developed. Active attacks form a special type of such privacy attacks, in which the adversary enrols a number of fake users, often called sybils, to the social network, allowing the adversary to create unique structural patterns later used to re-identify the sybil nodes and other users after anonymisation. Several studies have shown that adding a small amount of noise to the published graph already suffices to mitigate such active attacks. Consequently, active attacks have been dubbed a negligible threat to privacy-preserving social graph publication. In this paper, we argue that these studies unveil shortcomings of specific attacks, rather than inherent problems of active attacks as a general strategy. In order to support this claim, we develop the notion of a robust active attack, which is an active attack that is resilient to small perturbations of the social network graph. We formulate the design of robust active attacks as an optimisation problem and we give definitions of robustness for different stages of the active attack strategy. Moreover, we introduce various heuristics to achieve these notions of robustness and experimentally show that the new robust attacks are considerably more resilient than the original ones, while remaining at the same level of feasibility.

cs.SI↗

arXiv2017

Rethinking $(k,\ell)$-anonymity in social graphs: $(k,\ell)$-adjacency anonymity and $(k,\ell)$-(adjacency) anonymous transformations

This paper treats the privacy-preserving publication of social graphs in the presence of active adversaries, that is, adversaries with the ability to introduce sybil nodes in the graph prior to publication and leverage them to create unique fingerprints for a set of victim nodes and re-identify them after publication. Stemming from the notion of $(k,\ell)$-anonymity, we introduce $(k,\ell)$-anonymous transformations, characterising graph perturbation methods that ensure protection from active adversaries levaraging up to $\ell$ sybil nodes. Additionally, we introduce a new privacy property: $(k,\ell)$-adjacency anonymity, which relaxes the assumption made by $(k,\ell)$-anonymity that adversaries can control all distances between sybil nodes and the rest of the nodes in the graph. The new privacy property is in turn the basis for a new type of graph perturbation: $(k,\ell)$-adjacency anonymous transformations. We propose algorithms for obtaining $(k,1)$-adjacency anonymous transformations for arbitrary values of $k$, as well as $(2,\ell)$-adjacency anonymous transformations for small values of $\ell$.

cs.SI↗