An increasing number of countries implement Internet censorship at different scales and for a variety of reasons. Several censorship resistance systems (CRSs) have emerged to help bypass such blocks. The diversity of the censor’s attack landscape has led to an arms race, leading to a dramatic speed of evolution of CRSs. The inherent complexity of CRSs and the breadth of work in this area makes it hard to contextualize the censor’s capabilities and censorship resistance strategies. To address these challenges, we conducted a comprehensive survey of CRSs-deployed tools as well as those discussed in academic literature-to systematize censorship resistance systems by their threat model and corresponding defenses. To this end, we first sketch a comprehensive attack model to set out the censor’s capabilities, coupled with discussion on the scope of censorship, and the dynamics that influence the censor’s decision. Next, we present an evaluation framework to systematize censorship resistance systems by their security, privacy, performance and deployability properties, and show how these systems map to the attack model. We do this for each of the functional phases that we identify for censorship resistance systems: communication establishment, which involves distribution and retrieval of information necessary for a client to join the censorship resistance system; and conversation, where actual exchange of information takes place. Our evaluation leads us to identify gaps in the literature, question the assumptions at play, and explore possible mitigations.
Federated Learning (FL) is a data-minimization approach enabling collaborative model training across diverse clients with local data, avoiding direct data exchange. However, state-of-the-art FL solutions to identify fraudulent financial transactions exhibit a subset of the following limitations. They (1) lack a formal security definition and proof, (2) assume prior freezing of suspicious customers' accounts by financial institutions (limiting the solutions' adoption), (3) scale poorly, involving either $O(n^2)$ computationally expensive modular exponentiation (where $n$ is the total number of financial institutions) or highly inefficient fully homomorphic encryption, (4) assume the parties have already completed the identity alignment phase, hence excluding it from the implementation, performance evaluation, and security analysis, and (5) struggle to resist clients' dropouts. This work introduces Starlit, a novel scalable privacy-preserving FL mechanism that overcomes these limitations. It has various applications, such as enhancing financial fraud detection, mitigating terrorism, and enhancing digital health. We implemented Starlit and conducted a thorough performance analysis using synthetic data from a key player in global financial transactions. The evaluation indicates Starlit's scalability, efficiency, and accuracy.