Balázs Pejó

Martijn Oldenhof, Gergely Ács, Balázs Pejó et al.

To apply federated learning to drug discovery we developed a novel platform in the context of European Innovative Medicines Initiative (IMI) project MELLODDY (grant n°831472), which was comprised of 10 pharmaceutical companies, academic research labs, large industrial companies and startups. The MELLODDY platform was the first industry-scale platform to enable the creation of a global federated model for drug discovery without sharing the confidential data sets of the individual partners. The federated model was trained on the platform by aggregating the gradients of all contributing partners in a cryptographic, secure way following each training iteration. The platform was deployed on an Amazon Web Services (AWS) multi-account architecture running Kubernetes clusters in private subnets. Organisationally, the roles of the different partners were codified as different rights and permissions on the platform and administrated in a decentralized way. The MELLODDY platform generated new scientific discoveries which are described in a companion paper.

Balazs Pejo, Mina Remeli, Adam Arany et al.

Pharmaceutical industry can better leverage its data assets to virtualize drug discovery through a collaborative machine learning platform. On the other hand, there are non-negligible risks stemming from the unintended leakage of participants' training data, hence, it is essential for such a platform to be secure and privacy-preserving. This paper describes a privacy risk assessment for collaborative modeling in the preclinical phase of drug discovery to accelerate the selection of promising drug candidates. After a short taxonomy of state-of-the-art inference attacks we adopt and customize several to the underlying scenario. Finally we describe and experiments with a handful of relevant privacy protection techniques to mitigate such attacks.

Delio Jaramillo Velez, Gergely Biczok, Alexandre Graell i Amat et al.

Cross-silo federated learning allows multiple organizations to collaboratively train machine learning models without sharing raw data, but client updates can still leak sensitive information through inference attacks. Secure aggregation protects privacy by hiding individual updates, yet it complicates contribution evaluation, which is critical for fair rewards and detecting low-quality or malicious participants. Existing marginal-contribution methods, such as the Shapley value, are incompatible with secure aggregation, and practical alternatives, such as Leave-One-Out, are crude and rely on self-evaluation. We introduce two marginal-difference contribution scores compatible with secure aggregation. Fair-Private satisfies standard fairness axioms, while Everybody-Else eliminates self-evaluation and provides resistance to manipulation, addressing a largely overlooked vulnerability. We provide theoretical guarantees for fairness, privacy, robustness, and computational efficiency, and evaluate our methods on multiple medical image datasets and CIFAR10 in cross-silo settings. Our scores consistently outperform existing baselines, better approximate Shapley-induced client rankings, and improve downstream model performance as well as misbehavior detection. These results demonstrate that fairness, privacy, robustness, and practical utility can be achieved jointly in federated contribution evaluation, offering a principled solution for real-world cross-silo deployments.

Balazs Pejo, Nikolett Kapui

Almost 50 years after the invention of SQL, injection attacks are still top-tier vulnerabilities of today's ICT systems. Consequently, SQLi detection is still an active area of research, where the most recent works incorporate machine learning techniques into the proposed solutions. In this work, we highlight the shortcomings of the previous ML-based results focusing on four aspects: the evaluation methods, the optimization of the model parameters, the distribution of utilized datasets, and the feature selection. Since no single work explored all of these aspects satisfactorily, we fill this gap and provide an in-depth and comprehensive empirical analysis. Moreover, we cross-validate the trained models by using data from other distributions. This aspect of ML models (trained for SQLi detection) was never studied. Yet, the sensitivity of the model's performance to this is crucial for any real-life deployment. Finally, we validate our findings on a real-world industrial SQLi dataset.

Balazs Pejo

Federated learning offers a privacy-friendly collaborative learning framework, yet its success, like any joint venture, hinges on the contributions of its participants. Existing client evaluation methods predominantly focus on model performance, such as accuracy or loss, which represents only one dimension of a machine learning model's overall utility. In contrast, this work investigates the critical, yet overlooked, issue of client contributions towards a model's trustworthiness -- specifically, its reliability (tolerance to noisy data), resilience (resistance to adversarial examples), and fairness (measured via demographic parity). To quantify these multifaceted contributions, we employ the state-of-the-art approximation of the Shapley value, a principled method for value attribution. Our results reveal that no single client excels across all dimensions, which are largely independent from each other, highlighting a critical flaw in current evaluation scheme: no single metric is adequate for comprehensive evaluation and equitable rewarding allocation.

Marvin Xhemrishi, Alexandre Graell i Amat, Balázs Pejó

Federated learning with secure aggregation enables private and collaborative learning from decentralised data without leaking sensitive client information. However, secure aggregation also complicates the detection of malicious client behaviour and the evaluation of individual client contributions to the learning. To address these challenges, QI (Pejo et al.) and FedGT (Xhemrishi et al.) were proposed for contribution evaluation (CE) and misbehaviour detection (MD), respectively. QI, however, lacks adequate MD accuracy due to its reliance on the random selection of clients in each training round, while FedGT lacks the CE ability. In this work, we combine the strengths of QI and FedGT to achieve both robust MD and accurate CE. Our experiments demonstrate superior performance compared to using either method independently.

István András Seres, Balázs Pejó, Péter Burcsi

Fuzzy Message Detection (FMD) is a recent cryptographic primitive invented by Beck et al. (CCS'21) where an untrusted server performs coarse message filtering for its clients in a recipient-anonymous way. In FMD - besides the true positive messages - the clients download from the server their cover messages determined by their false-positive detection rates. What is more, within FMD, the server cannot distinguish between genuine and cover traffic. In this paper, we formally analyze the privacy guarantees of FMD from three different angles. First, we analyze three privacy provisions offered by FMD: recipient unlinkability, relationship anonymity, and temporal detection ambiguity. Second, we perform a differential privacy analysis and coin a relaxed definition to capture the privacy guarantees FMD yields. Finally, we simulate FMD on real-world communication data. Our theoretical and empirical results assist FMD users in adequately selecting their false-positive detection rates for various applications with given privacy requirements.

Mathias P. M. Parisot, Balazs Pejo, Dayana Spagnuelo

Machine learning models' goal is to make correct predictions for specific tasks by learning important properties and patterns from data. By doing so, there is a chance that the model learns properties that are unrelated to its primary task. Property Inference Attacks exploit this and aim to infer from a given model (\ie the target model) properties about the training dataset seemingly unrelated to the model's primary goal. If the training data is sensitive, such an attack could lead to privacy leakage. This paper investigates the influence of the target model's complexity on the accuracy of this type of attack, focusing on convolutional neural network classifiers. We perform attacks on models that are trained on facial images to predict whether someone's mouth is open. Our attacks' goal is to infer whether the training dataset is balanced gender-wise. Our findings reveal that the risk of a privacy breach is present independently of the target model's complexity: for all studied architectures, the attack's accuracy is clearly over the baseline. We discuss the implication of the property inference on personal data in the light of Data Protection Regulations and Guidelines.

Frederick Ayala-Gomez, Ismo Horppu, Erlin Gulbenkoglu et al.

Mobile app developers use paid advertising campaigns to acquire new users. Marketing managers decide where to spend and how much to spend based on the campaigns' performance. Apple's new privacy mechanisms have a profound impact on how performance marketing is measured. Starting iOS 14.5, all apps must get system permission for tracking explicitly via the new App Tracking Transparency Framework, which shows the users a pop-up asking if they give the app permission to track. If a user does not allow tracking, the required identifier to deterministically find the online advertising campaign that brought the user to install the app is not shared. Instead of relying on individual identifiers, Apple proposed a new performance mechanism called conversion value, which is an integer set by the apps for each user, and the developers can get the number of installs per conversion value for each campaign. However, interpreting how conversion values are used to measure the campaigns performance is not obvious because it requires a method to translate the conversion values to revenue. This paper investigates the task of attributing revenue to advertising campaigns using the reported conversion values per campaign. Our contributions are to formalize the problem, find the theoretically optimal revenue attribution function for any conversion value schema, and show empirical results on past data of a free-to-play mobile game using different conversion value schemas.

Balázs Pejó, Gergely Biczók

Federated learning algorithms are developed both for efficiency reasons and to ensure the privacy and confidentiality of personal and business data, respectively. Despite no data being shared explicitly, recent studies showed that the mechanism could still leak sensitive information. Hence, secure aggregation is utilized in many real-world scenarios to prevent attribution to specific participants. In this paper, we focus on the quality of individual training datasets and show that such quality information could be inferred and attributed to specific participants even when secure aggregation is applied. Specifically, through a series of image recognition experiments, we infer the relative quality ordering of participants. Moreover, we apply the inferred quality information to detect misbehaviours, to stabilize training performance, and to measure the individual contributions of participants.

Damien Desfontaines, Balázs Pejó

Shortly after it was first introduced in 2006, differential privacy became the flagship data privacy definition. Since then, numerous variants and extensions were proposed to adapt it to different scenarios and attacker models. In this work, we propose a systematic taxonomy of these variants and extensions. We list all data privacy definitions based on differential privacy, and partition them into seven categories, depending on which aspect of the original definition is modified. These categories act like dimensions: variants from the same category cannot be combined, but variants from different categories can be combined to form new definitions. We also establish a partial ordering of relative strength between these notions by summarizing existing results. Furthermore, we list which of these definitions satisfy some desirable properties, like composition, post-processing, and convexity by either providing a novel proof or collecting existing ones.

Balazs Pejo, Qiang Tang, Gergely Biczok

Machine learning algorithms have reached mainstream status and are widely deployed in many applications. The accuracy of such algorithms depends significantly on the size of the underlying training dataset; in reality a small or medium sized organization often does not have the necessary data to train a reasonably accurate model. For such organizations, a realistic solution is to train their machine learning models based on their joint dataset (which is a union of the individual ones). Unfortunately, privacy concerns prevent them from straightforwardly doing so. While a number of privacy-preserving solutions exist for collaborating organizations to securely aggregate the parameters in the process of training the models, we are not aware of any work that provides a rational framework for the participants to precisely balance the privacy loss and accuracy gain in their collaboration. In this paper, by focusing on a two-player setting, we model the collaborative training process as a two-player game where each player aims to achieve higher accuracy while preserving the privacy of its own dataset. We introduce the notion of Price of Privacy, a novel approach for measuring the impact of privacy protection on the accuracy in the proposed framework. Furthermore, we develop a game-theoretical model for different player types, and then either find or prove the existence of a Nash Equilibrium with regard to the strength of privacy protection for each player. Using recommendation systems as our main use case, we demonstrate how two players can make practical use of the proposed theoretical framework, including setting up the parameters and approximating the non-trivial Nash Equilibrium.