This paper presents a secure aggregation system Armadillo that has disruptive resistance against adversarial clients, such that any coalition of malicious clients can affect the aggregation result only by misreporting their private inputs in a pre-defined legitimate range. Armadillo is designed for federated learning setting, where a single powerful server interacts with many weak clients iteratively to train models on client's private data. While a few prior works consider disruption resistance under such setting, for an aggregation on n clients they either require high cost per client (Chowdhury et al. CCS '22) or concretely many rounds that is logarithmic in n (Bell et al. USENIX Security '23). Although disruption resistance can be achieved generically with zero-knowledge proof techniques (which we also use in this paper), we realize an efficient system with two new designs: 1) a simple two-layer secure aggregation protocol that requires only simple arithmetic computation; 2) an agreement protocol that removes the effect of malicious clients from the aggregation with low round complexity. With these techniques, Armadillo runs in 3 rounds per aggregation (our round complexity is independent of n) with computationally lightweight server and clients.

Federated Learning (FL) solutions with central Differential Privacy (DP) have seen large improvements in their utility in recent years arising from the matrix mechanism, while FL solutions with distributed (more private) DP have lagged behind. In this work, we introduce the distributed matrix mechanism to achieve the best-of-both-worlds; better privacy of distributed DP and better utility from the matrix mechanism. We accomplish this using a novel cryptographic protocol that securely transfers sensitive values across client committees of different training iterations with constant communication overhead. This protocol accommodates the dynamic participation of users required by FL, including those that may drop out from the computation. We provide experiments which show that our mechanism indeed significantly improves the utility of FL models compared to previous distributed DP mechanisms, with little added overhead.

As large language models (LLMs) become more powerful, the computation required to run these models is increasingly outsourced to a third-party cloud. While this saves clients’ computation, it risks leaking the clients’ LLM queries to the cloud provider. Fully homomorphic encryption (FHE) presents a natural solution to this problem: simply encrypt the query and evaluate the LLM homomorphically on the cloud machine. The result remains encrypted and can only be learned by the client who holds the secret key. In this work, we present a GPU-accelerated implementation of FHE and use this implementation to benchmark an encrypted GPT-2 forward pass, with runtimes over $200\times$ faster than the CPU baseline. We also present novel and extensive experimental analysis of approximations of LLM activation functions to maintain accuracy while achieving this performance.

The YOSO (You Only Speak Once) model, introduced by Gentry et al. (CRYPTO 2021), helps to achieve strong security guarantees in cryptographic protocols for large-scale distributed settings.

Fair decision-making in Machine Learning (ML) remains a critical challenge, particularly when access to sensitive information is restricted due to legal, ethical, or organizational constraints. These limitations affect both accuracy and fairness, creating tradeoffs central to the deployment of ML systems in the real world. While prior work has studied fairness-accuracy tradeoffs, most approaches focus on model outputs rather than directly examining how restricted data access impacts fairness. This leaves an important gap: understanding how fairness constraints affect model performance under real-world data restrictions . To address this gap, we propose a framework that explicitly models fairness-accuracy tradeoffs in data-restricted environments. Unlike prior work, our approach analyzes the behavior of the optimal Bayesian classifier using a discrete approximation of the data distribution, allowing us to systematically isolate the effects of fairness constraints. We evaluate our framework on three benchmark datasets—Adult, Law, and Dutch Census—revealing key insights: (1) enforcing equal accuracy on imbalanced datasets can substantially degrade performance under additional fairness constraints, (2) individual and group fairness often impose conflicting constraints, and (3) decorrelating sensitive attributes from features does not usually reduce accuracy. These findings demonstrate that our framework provides an effective, structured approach for practitioners to assess fairness constraints in decision-making pipelines.

In this work we study the efficiency of Zero-Knowledge (ZK) arguments of knowledge, particularly exploring Multi-Verifier ZK (MVZK) protocols as a midway point between Non-Interactive ZK and Designated-Verifier ZK, offering versatile applications across various domains. We introduce a new MVZK protocol designed for the preprocessing model, allowing any constant fraction of verifiers to be corrupted, potentially colluding with the prover. Our contributions include the first MVZK over rings. Unlike recent prior works on fields in the dishonest majority case, our protocol demonstrates communication complexity independent of the number of verifiers, contrasting the linear complexity of previous approaches. This key advancement ensures improved scalability and efficiency. We provide an end-to-end implementation of our protocol. The benchmark shows that it achieves a throughput of 1.47 million gates per second for 64 verifiers with 50% corruption, and 0.88 million gates per second with 75% corruption.

The growing use of machine learning (ML) has raised concerns that an ML model may reveal private information about an individual who has contributed to the training dataset. To prevent leakage of sensitive data, we consider using differentially- private (DP), synthetic training data instead of real training data to train an ML model. A key desirable property of synthetic data is its ability to preserve the low-order marginals of the original distribution. Our main contribution comprises novel upper and lower bounds on the excess empirical risk of linear models trained on such synthetic data, for continuous and Lipschitz loss functions. We perform extensive experimentation alongside our theoretical results.

Anonymous Zether, proposed by Bünz et al. (FC, 2020) and subsequently improved by Diamond (IEEE S&P, 2021) is an account-based confidential payment mechanism that works by using a smart contract to achieve privacy (i.e. identity of receivers to transactions and payloads are hidden). In this work, we look at simplifying the existing protocol while also achieving batching of transactions for multiple receivers, while ensuring consensus and forward secrecy. To the best of our knowledge, this work is the first to formally study the notion of forward secrecy in the setting of blockchain, borrowing a very popular and useful idea from the world of secure messaging. Specifically, we introduce:•FUL-Zether, a forward-secure version of Zether (Bünz et al. , FC, 2020),•PRIvate DEcentralized Confidential Transactions (PriDe CT), a much-simplified version of Anonymous Zether that achieves competitive performance and enables batching of transactions for multiple receivers.•PRIvate DEcentralized Forward-secure Until Last update Confidential Transactions (PriDeFUL CT), a forward-secure version of PriDe CT.We also present an open-source, Ethereum-based implementation of our system. PriDe CT uses linear homomor-phic encryption as Anonymous Zether but with simpler zero-knowledge proofs. PriDeFUL CT uses an updatable public key encryption scheme to achieve forward secrecy by introducing a new DDH-based construction in the standard model.In terms of transaction sizes, Quisquis (Asiacrypt, 2019), which is the only cryptocurrency that supports batchability (albeit in the UTXO model), has 15 times more group elements than PriDe CT. Meanwhile, for a ring of N receivers, Anonymous Zether requires 6 log N more terms even without accounting for the ability to batch in PriDe CT. Further, our implementation indicates that, for N = 32, even if there were 7 intended receivers, PriDe CT outperforms Anonymous Zether in proving time and gas consumption.

This work introduces MicroSecAgg, a framework that addresses the intricacies of secure aggregation in the single-server landscape, specifically tailored to situations where distributed trust among multiple non-colluding servers presents challenges. Our protocols are purpose-built to handle situations featuring multiple successive aggregation phases among a dynamic pool of clients who can drop out during the aggregation. Our different protocols thrive in three distinct cases: firstly, secure aggregation within a small input domain; secondly, secure aggregation within a large input domain; and finally, facilitating federated learning for the cases where moderately sized models are considered. Compared to the prior works of Bonawitz et al. (CCS 2017), Bell et al. (CCS 2020), and the recent work of Ma et al. (S&P 2023), our approach significantly reduces the overheads. In particular, MicroSecAgg halves the round complexity to just 3 rounds, thereby offering substantial improvements in communication cost efficiency. Notably, it outperforms Ma et al. by a factor of n on the user side, where n represents the number of users. Furthermore, in MicroSecAgg the computation complexity of each aggregation per user exhibits a logarithmic growth with respect to $n$, contrasting with the linearithmic or quadratic growth observed in Ma et al. and Bonawitz et al., respectively. We also require linear (in n) computation work from the server as opposed to quadratic in Bonawitz et al., or linearithmic in Ma et al. and Bell et al. In the realm of federated learning, a delicate tradeoff comes into play: our protocols shine brighter as the number of participating parties increases, yet they exhibit diminishing computational efficiency as the sheer volume of weights/parameters increases significantly. We report an implementation of our system and compare the performance against prior works, demonstrating that MicroSecAgg significantly reduces the computational burden and the message size.

Increases in the deployment of machine learning algorithms for applications that deal with sensitive data have brought attention to the issue of fairness in machine learning. Many works have been devoted to applications that require different demographic groups to be treated fairly. However, algorithms that aim to satisfy inter-group fairness (also called group fairness) may inadvertently treat individuals within the same demographic group unfairly. To address this issue, this article introduces a formal definition of within-group fairness that maintains fairness among individuals from within the same group. A pre-processing framework is proposed to meet both inter- and within-group fairness criteria with little compromise in performance. The framework maps the feature vectors of members from different groups to an inter-group fair canonical domain before feeding them into a scoring function. The mapping is constructed to preserve the relative relationship between the scores obtained from the unprocessed feature vectors of individuals from the same demographic group, guaranteeing within-group fairness. This framework has been applied to the Adult, COMPAS risk assessment, and Law School datasets, and its performance is demonstrated and compared with two regularization-based methods in achieving inter-group and within-group fairness.

This paper introduces Flamingo, a system for secure aggregation of data across a large set of clients. In secure aggregation, a server sums up the private inputs of clients and obtains the result without learning anything about the individual inputs beyond what is implied by the final sum. Flamingo focuses on the multi-round setting found in federated learning in which many consecutive summations (averages) of model weights are performed to derive a good model. Previous protocols, such as Bell et al. (CCS ’20), have been designed for a single round and are adapted to the federated learning setting by repeating the protocol multiple times. Flamingo eliminates the need for the per-round setup of previous protocols, and has a new lightweight dropout resilience protocol to ensure that if clients leave in the middle of a sum the server can still obtain a meaningful result. Furthermore, Flamingo introduces a new way to locally choose the so-called client neighborhood introduced by Bell et al. These techniques help Flamingo reduce the number of interactions between clients and the server, resulting in a significant reduction in the end-to-end runtime for a full training session over prior work.We implement and evaluate Flamingo and show that it can securely train a neural network on the (Extended) MNIST and CIFAR-100 datasets, and the model converges without a loss in accuracy, compared to a non-private federated learning system.

We introduce a secure data-independent priority queue which supports polylogarithmic-time insertion operations and constant-time deletions and read-front (aka peek) operations as opposed to the originally introduced queue by Toft (PODC '11). Moreover, we minimize the number of comparisons required to perform different operations on Toft's priority queue. Data-independent data structures—first identified explicitly by Toft, and further elaborated by Mitchell and Zimmerman (STACS '14)—serve the purpose of computing on encrypted data without executing branching code which can be used to avoid prohibitively expensive operations in secure computation applications. Focusing on the costly sorting operations, we show significant asymptotic improvements over prior privacy preserving dark pool applications. Dark pools are securities-trading venues which attain ad-hoc order privacy, by matching orders outside of publicly visible exchanges via the so-called dark pool operators. In this paper, we describe an efficient and secure dark pool (implementing a full continuous double auction) based on our new priority queue. Our construction's security guarantees are cryptographic based on secure multiparty computation (MPC), and do not require that the dark pool operators are trusted. Our construction improves upon the asymptotic efficiency attained by previous efforts. Existing cryptographic dark pools process new orders in time which grows linearly in the size of the standing order book; ours does so in polylogarithmic time. We describe a concrete implementation of our MPC protocol with malicious security in the honest majority setting. We also report benchmarks of our implementation and compare them to prior works. Our protocol reduces the total running time by several orders of magnitude over prior secure dark pool solutions.

We present a novel approach to honest majority secure multiparty computation in the preprocessing model with information theoretic security that achieves the best online communication complexity. The online phase of our protocol requires 12 elements in total per multiplication gate with circuit-dependent preprocessing, or 20 elements in total with circuit-independent preprocessing. Prior works achieved linear online communication complexity in n, the number of parties, with the best prior existing solution involving 1.5n elements per multiplication gate. Only one recent work packing [28] achieves constant online communication complexity, but the constants are large (108 elements for passive security, and twice that for active security). That said, our protocol offers a very efficient information theoretic online phase for any number of parties.

We revisit the problem of low-memory robust simulation of interactive protocols over noisy channels. Haeupler [FOCS 2014] considered robust simulation of two-party interactive protocols over oblivious, as well as adaptive, noisy channels. Since the simulation does not need to have fixed communication pattern, the achieved communication rates can circumvent the lower bound proved by Kol and Raz [STOC 2013]. However, a drawback of this approach is that each party needs to remember the whole history of the simulated transcript. In a subsequent manuscript, Haeupler and Resch considered low-memory simulation. The idea was to view the original protocol as a computational DAG and only the identities of the nodes are saved (as opposed to the whole transcript history) for backtracking to reduce memory usage. In this paper, we consider low-memory robust simulation of more general client-server interactive protocols, in which a leader communicates with other members/servers, who do not communicate among themselves; this setting can be applied to information-theoretic multi-server Private Information Retrieval (PIR) schemes. We propose an information-theoretic technique that converts any correct PIR protocol that assumes reliable channels, into a protocol which is both correct and private in the presence of a noisy channel while keeping the space complexity to a minimum. Despite the huge attention that PIR protocols have received in the literature, the existing works assume that the parties communicate using noiseless channels. Moreover, we observe that the approach of Haeupler and Resch to just save the nodes in the aforementioned DAG without taking the transcript history into account will lead to a correctness issue even for oblivious corruptions. We resolve this issue by saving hashes of prefixes of past transcripts. Departing from the DAG representation also allows us to accommodate scenarios where a party can simulate its part of the protocol without any extra knowledge (such as the DAG representation of the whole protocol). In the the two-party setting, our simulation has the same dependence on the error rate as in the work of Haeupler, and in the client-server setting it also depends on the number of servers. Furthermore, since our approach does not remember the complete transcript history, our current technique can defend only against oblivious corruptions.