Privacy Preservation in Trust-Deficient Decentralized Systems
Résumé
Decentralized distributed systems can provide certain advantages over their centralized counterparts. These include improved fault tolerance and attack resistance due to the elimination of a single point of failure, better censorship resistance and openness because of the absence of a central authority, as well as higher autonomy of the nodes due to self-governance of resources. Distributed ledger and blockchain technologies have played a key role in the increasing adoption of decentralized systems. One of the notable contributions of blockchains to decentralized systems is that they are able to ensure properties such as integrity, immutability, and verifiability even in trust-deficient environments where nodes lack trust in each other. However, ensuring the privacy of users still remains a challenge in decentralized systems where the application requires users to contribute confidential or identifiable information. Privacy is particularly challenging to achieve in trust-deficient environments due to the nodes not being able to rely on fellow nodes for privacy guarantees. My work addresses privacy preservation in trust-deficient decentralized systems. In this habilitation thesis, we include some selected contributions from our work in three broad areas: privacy-preserving decentralized reputation systems, privacy-preserving message routing in decentralized networks, and privacy preservation in decentralized financial networks. The reputation of a user in a distributed system may be computed as an aggregate of the feedback provided by fellow users. Privacy-preserving reputation systems enable users to provide feedback in a private and thus uninhibited manner. We present several contributions in this area, which include a decentralized privacy-preserving reputation protocol based on the computation of mean that is secure under the malicious adversarial model. The protocol offers significant improvement in performance as compared to earlier protocols in the literature. Additionally, we define an attack called Reputation-based Re-identification (RR attack), which can link successive contributions provided by a participant in participatory sensing applications and subsequently re-identify them. We then propose PrivaSense, a privacy-preserving reputation system that defends against this attack. Experiments on a real dataset demonstrate that PrivaSense is successful in decreasing the number of contributions linked to their providers. Furthermore, we propose a voting protocol (as a generalized instance of a privacy-preserving reputation protocol) that ensures transparency, confidentiality, and integrity in a decentralized trust-deficient setup. The persistence and immutability of the protocol's communication allow verifiability of the outcome by the voters themselves. Our contribution on secure voting is further extended by our work on collusion-resistant worker set selection. We propose protocols that select a subset of workers (who process data during the protocol execution) from the set of participants such that the risk of collusion between workers is minimized. Mobile Delay Tolerant Networks (MDTNs) are composed of mobile devices that communicate in a decentralized manner without the help of fixed infrastructure. A prediction-based routing protocol for MDTNs functions by forwarding a message from one intermediate node to another if the latter has higher probability of encountering the destination node. However, this process compromises the privacy of the nodes by revealing their mobility patterns. We propose the Privacy-Preserving Probabilistic Prediction-based Routing (4PR) protocol that forwards messages by comparing information about communities of nodes instead of individual nodes in order to protect their privacy. The protocol computes a probability function in a decentralized privacy-preserving manner. Simulations on a community-based mobility model demonstrate that our protocol is able to preserve privacy while offering performance comparable to protocols that do not protect privacy. Many Decentralized Finance (DeFi) Peer-to-Peer (P2P) lending platforms offer users to obtain a loan by committing a collateral or by calculating a credit score. However, the requirements of collateral and credit history are quite burdensome for certain groups of users. We propose to use a social trustworthiness score drawn from users' social interactions as an alternative risk mitigator for lending instead of collateral. Privacy considerations are taken into account in order to protect the borrower's privacy despite the use of social interaction data. Another application that we address is Supplier Impersonation Fraud (SIF) detection in the Business-to-Business (B2B) context. This type of fraud occurs when a company supplying goods and services to another company is impersonated by a fraudster in order to trigger a payment to an illegitimate bank account. We introduce GraphSIF, a SIF detection system that aims to infer knowledge from the relational properties created by the transactions of a company. GraphSIF analyzes data that has been heavily anonymized in order to preserve the privacy of the participating companies. The classification of a targeted transaction is performed by first clustering the graphs, and then comparing the similarity between the targeted transaction's graph with the other graphs of its cluster. The model shows good efficiency in terms of computational time needed to create the behavior sequence and to classify the transactions.