Although distributed machine learning (distributed ML) is gaining considerable attention in the community, prior works have independently looked at instances of distributed ML in either the training or the inference phase. No prior work has examined the combined robustness stemming from distributing both the learning and the inference process. In this work, we explore, for the first time, the robustness of distributed ML models that are fully heterogeneous in training data, architecture, scheduler, optimizer, and other model parameters. Supported by theory and extensive experimental validation using CIFAR10 and FashionMNIST, we show that such properly distributed ML instantiations achieve across-the-board improvements in accuracy-robustness tradeoffs against state-of-the-art transfer-based attacks that could otherwise not be realized by current ensemble or federated learning instantiations. For instance, our experiments on CIFAR10 show that for the Common Weakness attack, one of the most powerful state-of-the-art transfer-based attacks, our method improves robust accuracy by up to 40%, with a minimal impact on clean task accuracy.

Although promising, existing defenses against query-based attacks share a common limitation: they offer increased robustness against attacks at the price of a considerable accuracy drop on clean samples. In this work, we show how to efficiently establish, at test-time, a solid tradeoff between robustness and accuracy when mitigating query-based attacks. Given that these attacks necessarily explore low-confidence regions, our insight is that activating dedicated defenses, such as random noise defense and random image transformations, only for low-confidence inputs is sufficient to prevent them. Our approach is independent of training and supported by theory. We verify the effectiveness of our approach for various existing defenses by conducting extensive experiments on CIFAR-10, CIFAR-100, and ImageNet. Our results confirm that our proposal can indeed enhance these defenses by providing better tradeoffs between robustness and accuracy when compared to state-of-the-art approaches while being completely training-free.

Solana gained considerable attention as one of the most popular blockchain platforms for deploying decentralized applications. Compared to Ethereum, however, we observe a lack of research on how Solana smart contract developers handle security, what challenges they encounter, and how this affects the overall security of the ecosystem. To address this, we conducted the first comprehensive study on the Solana platform consisting of a 90-minute Solana smart contract code review task with 35 participants followed by interviews with a subset of seven participants. Our study shows, quite alarmingly, that none of the participants could detect all important security vulnerabilities in a code review task and that 83% of the participants are likely to release vulnerable smart contracts. Our study also sheds light on the root causes of developers' challenges with Solana smart contract development, suggesting the need for better security guidance and resources. In spite of these challenges, our automated analysis on currently deployed Solana smart contracts surprisingly suggests that the prevalence of vulnerabilities - especially those pointed out as the most challenging in our developer study - is below 0.3%. We explore the causes of this counter-intuitive resilience and show that frameworks, such as Anchor, are aiding Solana developers in deploying secure contracts.

Extensive research on Nakamoto-style consensus protocols has shown that network delays degrade the security of these protocols. Established results indicate that, perhaps surprisingly, maximal security is achieved when the network is as small as two nodes due to increased delays in larger networks. This contradicts the very foundation of blockchains, namely that decentralization improves security.In this paper, we take a closer look at how the network scale affects security of Nakamoto-style blockchains. We argue that a crucial aspect has been neglected in existing security models: the larger the network, the harder it is for an attacker to control a significant amount of power. To this end, we introduce a probabilistic corruption model to express the increasing difficulty for an attacker to corrupt resources in larger networks. Based on our model, we analyze the impact of the number of nodes on the (maximum) network delay and the fraction of adversarial power. In particular, we show that (1) increasing the number of nodes eventually violates security, but (2) relying on a small number of nodes does not provide decent security provisions either. We then validate our analysis by means of an empirical evaluation emulating hundreds of thousands of nodes in deployments such as Bitcoin, Monero, Cardano, and Ethereum Classic. Based on our empirical analysis, we concretely analyze the impact of various real-world parameters and configurations on the consistency bounds in existing deployments and on the adversarial power that can be tolerated while providing security. As far as we are aware, this is the first work that analytically and empirically explores the real-world tradeoffs achieved by current popular Nakamoto-style deployments.

Proof of Stake (PoS) blockchain protocols emerged as a promising alternative to the largely energy-wasteful proof of work mechanisms currently in place. In contrast to computing power, however, “stake” is a virtual resource that can be replicated or reused, opening the door to attack vectors that have no counterpart in a PoW setting, and are much harder to defeat. We present PoTS (Proof of TEE-Stake), a novel PoS protocol that leverages properties of trusted execution environments (TEEs) to limit the attack surface of malicious validators, and employs techniques such as forward security to guarantee protection against posterior-corruption attacks. We show that PoTS is secure against nothing at stake, grinding, and long range attacks down to realistic hardware assumptions on TEE and well-established cryptographic assumptions, and retains reasonable security even in face of compromised TEEs. We evaluate the performance of our proposal by means of implementation. Our evaluation results demonstrate that PoTS offers an excellent trade-off between security and performance.

Scalability remains one of the biggest challenges to the adoption of permissioned blockchain technologies for large-scale deployments. Namely, permissioned blockchains typically exhibit low latencies, compared to permissionless deployments—however at the cost of poor scalability. As a remedy, various solutions were proposed to capture “the best of both worlds”, targeting low latency and high scalability simultaneously. Among these, blockchain sharding emerges as the most prominent technique. Most existing sharding proposals exploit features of the permissionless model and are therefore restricted to cryptocurrency applications. A few permissioned sharding proposals exist, however, they either make strong trust assumptions on the number of faulty nodes or rely on trusted hardware or assume a static participation model where all nodes are expected to be available all the time. In practice, nodes may join and leave the system dynamically, which makes it challenging to establish how to shard and when.

Recent studies have shown that federated learning (FL) is vulnerable to poisoning attacks that inject a backdoor into the global model. These attacks are effective even when performed by a single client, and undetectable by most existing defensive techniques. In this paper, we propose Backdoor detection via Feedback-based Federated Learning (BAFFLE), a novel defense to secure FL against backdoor attacks. The core idea behind BAFFLE is to leverage data of multiple clients not only for training but also for uncovering model poisoning. We exploit the availability of diverse datasets at the various clients by incorporating a feedback loop into the FL process, to integrate the views of those clients when deciding whether a given model update is genuine or not. We show that this powerful construct can achieve very high detection rates against state-of-the-art backdoor attacks, even when relying on straightforward methods to validate the model. Through empirical evaluation using the CIFAR-10 and FEMNIST datasets, we show that by combining the feedback loop with a method that suspects poisoning attempts by assessing the per-class classification performance of the updated model, BAFFLE reliably detects state-of-the-art backdoor attacks with a detection accuracy of 100% and a false-positive rate below 5%. Moreover, we show that our solution can detect adaptive attacks aimed at bypassing the defense.