This is an extended version of our work in [16]. In this paper, we introduce two novel algorithms to collaboratively train Naive Bayes models across multiple private data sources: Federated Naive Bayes and Gossip Naive Bayes. Instead of directly providing access to their data, the data owners compute local updates that are then aggregated to build a global model. In order to also prevent indirect privacy leaks from the updates or from the final model, our algorithms protect the exchanged information with differential privacy. We experimentally evaluate our proposed approaches, examining different scenarios and focusing on potential real-world issues, such as different data owner offering different amounts of data or requesting different levels of privacy. Our results show that both Federated and Gossip Naive Bayes achieve similar accuracy to a “vanilla” Naive Bayes while maintaining reasonable privacy guarantees, while being extremely robust to heterogeneous data owners.

Current Machine Learning (ML) approaches typically present either a centralized or federated architecture. However, these architectures cannot easily keep up with some of the challenges introduced by recent trends, such as the growth in the number of IoT devices, increasing awareness about the privacy and security implications of extensive data collection, and the rise of graph-structured data and Graph Representation Learning. Systems based on either direct data collection or Federated Learning contain centralized, privileged systems that may act as scalability bottlenecks and dangerous single points of failure, while requiring users to trust the privacy protections and security practices in place. The combination of these issues ultimately leads to data waste, as opportunities to extract insights from available data are missed and thus the full societal benefits of advanced data analytics and ML are not realized.

In this thesis, we argue for a paradigm shift towards a completely decentralized and trustless architecture for privacy-aware Graph Representation Learning, which employs Gossip Learning and other gossip-based peer-to-peer techniques to achieve high levels of scalability and resilience while reducing the risk of privacy leaks. We then identify and pursue three key research directions necessary to achieve our vision: lifting unrealistic assumptions on Gossip Learning, identifying and developing specific use cases that are enabled or improved by gossip-based decentralization, and overcoming the obstacles to the deployment of decentralized training and inference for Graph Representation Learning models.

 Based on these key directions, our contributions are as follows. First, we analyze the robustness of Gossip Learning when several unrealistic but often assumed conditions are lifted. Then, we exploit Gossip Learning and gossip-based peer-to-peer protocols more in general across three use cases: the collaborative training of differentially-private Naive Bayes classifiers across organizations holding sensitive user data; the construction of decentralized, privacy-preserving data marketplaces; and the development and decentralization of early-stage IoT botnet detection systems based on Graph Representation Learning. Finally, we introduce a general framework for the fully-decentralized training of Graph Neural Networks, overcoming the typical requirement of these models to access non-local information during training and inference.

 The combination of these contributions removes major roadblocks towards decentralized graph learning, and also opens a new research direction aimed at further developing and optimizing the fully-decentralized training of Graph Representation Learning models.

Dagens metoder för maskininlärning (ML) har vanligtvis antingen en centraliserad eller federerad arkitektur. Dessa arkitekturer kan dock inte lätt hålla jämna steg med några av de utmaningar som introducerats av de senaste trenderna, som till exempel ökningen av antalet IoT-enheter, ökad medvetenhet om integritets- och säkerhetskonsekvenserna av omfattande datainsamling samt ökningen av grafstrukturerad data och Graph Representation Learning. System baserade på antingen direkt datainsamling eller federerad inlärning innehåller centraliserade, privilegierade system som kan vara flaskhalsar och riskerar bli kritiska sårbarhetspunkter. Samtidigt måste användarna lita på integritetsskyddet och säkerhetspraxis som finns. Kombinationen av dessa problem leder i slutändan till ett ineffektivt nyttjande av data, eftersom möjligheter att utvinna insikter från tillgänglig data inte utnyttjas och därmed inte realiserar de fulla samhällsnyttorna som är möjliga med avancerad dataanalys och ML.

I denna avhandling argumenterar vi för ett paradigmskifte mot en helt decentraliserad och tillitslös arkitektur för integritetsmedveten Graph Representation Learning, som använder Gossip Learning och andra gossip-baserade peer-to-peer-tekniker för att uppnå höga nivåer av skalbarhet och motståndskraft, samtidigt som den minskar risken för integritetsläckor. Vi identifierar och driver sedan tre viktiga forskningsinriktningar som är nödvändiga för att uppnå vår vision; att lyfta orealistiska antaganden om Gossip Learning, identifiera och utveckla specifika användningsfall som möjliggörs eller förbättras av gossip-baserad decentralisering, samt övervinna hindren för utplacering av decentraliserad utbildning och inferens för Graph Representation Learning modeller.

Baserat på dessa nyckelriktlinjer våra bidrag är följande. Först analyserar vi robustheten i Gossip Learning när flera orealistiska men ofta antagna villkor upphävs. Vi utnyttjar sedan Gossip Learning och gossip-baserade peer-to-peer-protokoll mer generellt i tre användningsfall: kollaborativ inlärning av differentiellt privata Naive Bayes-klassificerare över entiteter med känslig användardata; byggandet av decentraliserade datamarknadsplatser som bevarar integriteten; samt utveckling och decentralisering av IoT-botnätdetekterings\-system i ett tidigt skede baserade på Graph Representation Learning. Slutligen introducerar vi ett allmänt ramverk för helt decentraliserad utbildning av Graph Neural Networks, som eliminerar de typiska kraven för dessa modeller för att få tillgång till icke-lokal information under träning och inferens.

Kombinationen av dessa bidrag tar bort stora hinder mot decentraliserad grafinlärning, och öppnar också en ny forskningsriktning som syftar till att vidareutveckla och optimera den helt decentraliserade utbildningen av Graph Representation Learning modeller.

Growing privacy concerns regarding personal data disclosure are contrasting with the constant need of such information for data-driven applications. To address this issue, the combination of federated learning and differential privacy is now well-established in the domain of machine learning. These techniques allow to train deep neural networks without collecting the data and while preventing information leakage. However, there are many scenarios where simpler and more robust machine learning models are preferable. In this paper, we present a federated and differentially-private version of the Naive Bayes algorithm for classification. Our results show that, without data collection, the same performance of a centralized solution can be achieved on any dataset with only a slight increase in the privacy budget. Furthermore, if certain conditions are met, our federated solution can outperform a centralized approach.

Heterogeneous Information Network (HIN) embedding has been a prevalent approach to learn representations off semantically-rich heterogeneous networks. Most HIN embedding methods exploit meta-paths to retain high-order structures, yet, their performance is conditioned on the quality of the (generated/manually-defined) meta-paths and their suitability for the specific label set. Whereas other methods adjust random walks to harness or skip certain heterogeneous structures (e.g. node type(s)), in doing so, the adjusted random walker may casually omit other node/edge types. Our key insight is with no domain knowledge, the random walker should hold no assumptions about heterogeneous structure (i.e. edge types). Thus, aiming for a flexible and general method, we utilize network schema as a unique blueprint of HIN, and propose SchemaWalk, a random walk to uniformly sample all edge types within the network schema. Moreover, we identify the starvation phenomenon which induces random walkers on HINs to under- or over-sample certain edge types. Accordingly, we design SchemaWalkHO to skip local deficient connectivity to preserve uniform sampling distribution. Finally, we carry out node classification experiments on four real-world HINs, and provide in-depth qualitative analysis. The results highlight the robustness of our method regardless to the graph structure in contrast with the state-of-the-art baselines. 

One big challenge for Industry 4.0 is leveraging the large amount of data that remain unused after collection. A variety of commercial data marketplaces have emerged in recent years to tackle this task. Despite their different business models and target markets, such marketplaces share a number of common issues that slow the growth of the industry, including data discovery, transparency, data privacy and data valuation. Many academic designs have been proposed to address these issues, yet most of them remain unimplemented, due to complexity or inefficiency.

We argue that these issues can be addressed with a combination of blockchain-based infrastructure, privacy-preserving computing and machine learning-based valuation metrics. Furthermore, we discuss key enabling technologies in each of these areas that are feasible to deploy at scale and could thus be implemented in real-world marketplaces in the near future. We select such technologies based on their current maturity and their industrial prominence.

One big challenge for Industry 4.0 is leveraging the large amount of data that remain unused after collection. A variety of commercial data marketplaces have emerged in recent years to tackle this task. Despite their different business models and target markets, such marketplaces share a number of common issues that slow the growth of the industry, including data discovery, transparency, data privacy and data valuation. Many academic designs have been proposed to address these issues, yet most of them remain unimplemented, due to complexity or inefficiency. We argue that these issues can be addressed with a combination of blockchain-based infrastructure, privacy-preserving computing and machine learning-based valuation metrics. Furthermore, we discuss key enabling technologies in each of these areas that are feasible to deploy at scale and could thus be implemented in real-world marketplaces in the near future. We select such technologies based on their current maturity and their industrial prominence.

Advanced NLP models require huge amounts of data from various domains to produce high-quality representations. It is useful then for a few large public and private organizations to join their corpora during training. However, factors such as legislation and user emphasis on data privacy may prevent centralized orchestration and data sharing among these organizations. Therefore, for this specific scenario, we investigate how gossip learning, a massively-parallel, data-private, decentralized protocol, compares to a shared-dataset solution. We find that the application of Word2Vec in a gossip learning framework is viable. Without any tuning, the results are comparable to a traditional centralized setting, with a reduction in ground-truth similarity scores as low as 4.3%. Furthermore, the results are up to 54.8% better than independent local training.

IoT devices have been growing exponentially in the last few years. This growth makes them an attractive target for attackers due to their low computational power and limited security features. Attackers use IoT botnets as an instrument to perform DDoS attacks which caused major disruptions of Internet services in the last decade. While many works have tackled the task of detecting botnet attacks, only a few have considered early-stage detection of these botnets during their propagation phase.

While previous approaches analyze each network packet individually to predict its maliciousness, we propose a novel deep learning model called LiMNet (Lightweight Memory Network), which uses an internal memory component to capture the behaviour of each IoT device over time. This memory incorporates both packet features and behaviour of the peer devices. With this information, LiMNet achieves almost maximum AUROC classification scores, between 98.8% and 99.7%, with a 14% improvement over state of the art. LiMNet is also lightweight, performing inference almost 8 times faster than previous approaches.

We envision PDS², a decentralized data marketplace in which consumers submit their tasks to be run within the platform, on the data of willing providers. The goal of PDS2is to ensure that users maintain full control on their data and do not compromise their privacy, while being rewarded for the value that their data generates. In order to achieve this, our marketplace architecture employs blockchain technology, privacy-preserving computation and decentralized machine learning.

We then compare different potential solutions and identify the Ethereum blockchain, trusted execution environments and gossip learning as the most suitable for the implementation of PDS². We also discuss the main open challenges that are left to tackle and possible directions for future work

We envision PDS2, a decentralized data marketplace in which consumers submit their tasks to be run within the platform, on the data of willing providers. The goal of PDS2 is to ensure that users maintain full control on their data and do not compromise their privacy, while being rewarded for the value that their data generates. In order to achieve this, our marketplace architecture employs blockchain technology, privacypreserving computation and decentralized machine learning. We then compare different potential solutions and identify the Ethereum blockchain, trusted execution environments and gossip learning as the most suitable for the implementation of PDS2. We also discuss the main open challenges that are left to tackle and possible directions for future work.