Search

Machine unlearning, the process of selectively removing data from trained models, is increasingly crucial for addressing privacy concerns and knowledge gaps post-deployment. Despite this importance, existing approaches are often heuristic and lack formal guarantees. In this paper, we analyze the fundamental utility, time, and space complexity trade-offs of approximate unlearning, providing rigorous certification analogous to differential privacy. For in-distribution forget data -- data similar to the retain set -- we show that a surprisingly simple and general procedure, empirical risk minimization with output perturbation, achieves tight unlearning-utility-complexity trade-offs, addressing a previous theoretical gap on the separation from unlearning "for free" via differential privacy, which inherently facilitates the removal of such data. However, such techniques fail with out-of-distribution forget data -- data significantly different from the retain set -- where unlearning time complexity can exceed that of retraining, even for a single sample. To address this, we propose a new robust and noisy gradient descent variant that provably amortizes unlearning time complexity without compromising utility.

Federated learning (FL) is an appealing approach to training machine learning models without sharing raw data. However, standard FL algorithms are iterative and thus induce a significant communication cost. One-shot federated learning (OFL) trades the iterative exchange of models between clients and the server with a single round of communication, thereby saving substantially on communication costs. Not surprisingly, OFL exhibits a performance gap in terms of accuracy with respect to FL, especially under high data heterogeneity. We introduce FENS, a novel federated ensembling scheme that approaches the accuracy of FL with the communication efficiency of OFL. Learning in FENS proceeds in two phases: first, clients train models locally and send them to the server, similar to OFL; second, clients collaboratively train a lightweight prediction aggregator model using FL. We showcase the effectiveness of FENS through exhaustive experiments spanning several datasets and heterogeneity levels. In the particular case of heterogeneously distributed CIFAR-10 dataset, FENS achieves up to a 26.9% higher accuracy over state-of-the-art (SOTA) OFL, being only 3.1% lower than FL. At the same time, FENS incurs at most 4.3x more communication than OFL, whereas FL is at least 10.9x more communication-intensive than FENS., Comment: Accepted at NeurIPS 2024

Federated learning (FL) is an appealing paradigm that allows a group of machines (a.k.a. clients) to learn collectively while keeping their data local. However, due to the heterogeneity between the clients' data distributions, the model obtained through the use of FL algorithms may perform poorly on some client's data. Personalization addresses this issue by enabling each client to have a different model tailored to their own data while simultaneously benefiting from the other clients' data. We consider an FL setting where some clients can be adversarial, and we derive conditions under which full collaboration fails. Specifically, we analyze the generalization performance of an interpolated personalized FL framework in the presence of adversarial clients, and we precisely characterize situations when full collaboration performs strictly worse than fine-tuned personalization. Our analysis determines how much we should scale down the level of collaboration, according to data heterogeneity and the tolerable fraction of adversarial clients. We support our findings with empirical results on mean estimation and binary classification problems, considering synthetic and benchmark image classification datasets.

Memory disaggregation is an emerging data center architecture that improves resource utilization and scalability. Replication is key to ensure the fault tolerance of applications, but replicating shared data in disaggregated memory is hard. We propose SWARM (Swift WAit-free Replication in disaggregated Memory), the first replication scheme for in-disaggregated-memory shared objects to provide (1) single-roundtrip reads and writes in the common case, (2) strong consistency (linearizability), and (3) strong liveness (wait-freedom). SWARM makes two independent contributions. The first is Safe-Guess, a novel wait-free replication protocol with single-roundtrip operations. The second is In-n-Out, a novel technique to provide conditional atomic update and atomic retrieval of large buffers in disaggregated memory in one roundtrip. Using SWARM, we build SWARM-KV, a low-latency, strongly consistent and highly available disaggregated key-value store. We evaluate SWARM-KV and find that it has marginal latency overhead compared to an unreplicated key-value store, and that it offers much lower latency and better availability than FUSEE, a state-of-the-art replicated disaggregated key-value store., Comment: To appear in the proceedings of SOSP '24

Blockchain promises to make online services more fault tolerant due to their inherent distributed nature. Their ability to execute arbitrary programs in different geo-distributed regions and on diverse operating systems make them an alternative of choice to our dependence on unique software whose recent failure affected 8.5 millions of machines. As of today, it remains, however, unclear whether blockchains can truly tolerate failures. In this paper, we assess the fault tolerance of blockchain. To this end, we inject failures in controlled deployments of five modern blockchain systems, namely Algorand, Aptos, Avalanche, Redbelly and Solana. We introduce a novel sensitivity metric, interesting in its own right, as the difference between the integrals of two cumulative distribution functions, one obtained in a baseline environment and one obtained in an adversarial environment. Our results indicate that (i) all blockchains except Redbelly are highly impacted by the failure of a small part of their network, (ii) Avalanche and Redbelly benefit from the redundant information needed for Byzantine fault tolerance while others are hampered by it, and more dramatically (iii) Avalanche and Solana cannot recover from localised transient failures., Comment: 16 pages, 6 figures

AI assistants are being increasingly used by students enrolled in higher education institutions. While these tools provide opportunities for improved teaching and education, they also pose significant challenges for assessment and learning outcomes. We conceptualize these challenges through the lens of vulnerability, the potential for university assessments and learning outcomes to be impacted by student use of generative AI. We investigate the potential scale of this vulnerability by measuring the degree to which AI assistants can complete assessment questions in standard university-level STEM courses. Specifically, we compile a novel dataset of textual assessment questions from 50 courses at EPFL and evaluate whether two AI assistants, GPT-3.5 and GPT-4 can adequately answer these questions. We use eight prompting strategies to produce responses and find that GPT-4 answers an average of 65.8% of questions correctly, and can even produce the correct answer across at least one prompting strategy for 85.1% of questions. When grouping courses in our dataset by degree program, these systems already pass non-project assessments of large numbers of core courses in various degree programs, posing risks to higher education accreditation that will be amplified as these models improve. Our results call for revising program-level assessment design in higher education in light of advances in generative AI., Comment: 20 pages, 8 figures

The ability of a peer-to-peer (P2P) system to effectively host decentralized applications often relies on the availability of a peer-sampling service, which provides each participant with a random sample of other peers. Despite the practical effectiveness of existing peer samplers, their ability to produce random samples within a reasonable time frame remains poorly understood from a theoretical standpoint. This paper contributes to bridging this gap by introducing PeerSwap, a peer-sampling protocol with provable randomness guarantees. We establish execution time bounds for PeerSwap, demonstrating its ability to scale effectively with the network size. We prove that PeerSwap maintains the fixed structure of the communication graph while allowing sequential peer position swaps within this graph. We do so by showing that PeerSwap is a specific instance of an interchange process, a renowned model for particle movement analysis. Leveraging this mapping, we derive execution time bounds, expressed as a function of the network size N. Depending on the network structure, this time can be as low as a polylogarithmic function of N, highlighting the efficiency of PeerSwap. We implement PeerSwap and conduct numerical evaluations using regular graphs with varying connectivity and containing up to 32768 (2^15) peers. Our evaluation demonstrates that PeerSwap quickly provides peers with uniform random samples of other peers.

Data centers increasingly host mutually distrustful users on shared infrastructure. A powerful tool to safeguard such users are digital signatures. Digital signatures have revolutionized Internet-scale applications, but current signatures are too slow for the growing genre of microsecond-scale systems in modern data centers. We propose DSig, the first digital signature system to achieve single-digit microsecond latency to sign, transmit, and verify signatures in data center systems. DSig is based on the observation that, in many data center applications, the signer of a message knows most of the time who will verify its signature. We introduce a new hybrid signature scheme that combines cheap single-use hash-based signatures verified in the foreground with traditional signatures pre-verified in the background. Compared to prior state-of-the-art signatures, DSig reduces signing time from 18.9 to 0.7 us and verification time from 35.6 to 5.1 us, while keeping signature transmission time below 2.5 us. Moreover, DSig achieves 2.5x higher signing throughput and 6.9x higher verification throughput than the state of the art. We use DSig to (a) bring auditability to two key-value stores (HERD and Redis) and a financial trading system (based on Liquibook) for 86% lower added latency than the state of the art, and (b) replace signatures in BFT broadcast and BFT replication, reducing their latency by 73% and 69%, respectively, Comment: To appear in the proceedings of OSDI '24. Authors listed in alphabetical order

Batch normalization has proven to be a very beneficial mechanism to accelerate the training and improve the accuracy of deep neural networks in centralized environments. Yet, the scheme faces significant challenges in federated learning, especially under high data heterogeneity. Essentially, the main challenges arise from external covariate shifts and inconsistent statistics across clients. We introduce in this paper Federated BatchNorm (FBN), a novel scheme that restores the benefits of batch normalization in federated learning. Essentially, FBN ensures that the batch normalization during training is consistent with what would be achieved in a centralized execution, hence preserving the distribution of the data, and providing running statistics that accurately approximate the global statistics. FBN thereby reduces the external covariate shift and matches the evaluation performance of the centralized setting. We also show that, with a slight increase in complexity, we can robustify FBN to mitigate erroneous statistics and potentially adversarial attacks.

Byzantine-resilient distributed machine learning seeks to achieve robust learning performance in the presence of misbehaving or adversarial workers. While state-of-the-art (SOTA) robust distributed gradient descent (Robust-DGD) methods were proven theoretically optimal, their empirical success has often relied on pre-aggregation gradient clipping. However, the currently considered static clipping strategy exhibits mixed results: improving robustness against some attacks while being ineffective or detrimental against others. We address this gap by proposing a principled adaptive clipping strategy, termed Adaptive Robust Clipping (ARC). We show that ARC consistently enhances the empirical robustness of SOTA Robust-DGD methods, while preserving the theoretical robustness guarantees. Our analysis shows that ARC provably improves the asymptotic convergence guarantee of Robust-DGD in the case when the model is well-initialized. We validate this theoretical insight through an exhaustive set of experiments on benchmark image classification tasks. We observe that the improvement induced by ARC is more pronounced in highly heterogeneous and adversarial settings.

Decentralized learning is appealing as it enables the scalable usage of large amounts of distributed data and resources (without resorting to any central entity), while promoting privacy since every user minimizes the direct exposure of their data. Yet, without additional precautions, curious users can still leverage models obtained from their peers to violate privacy. In this paper, we propose Decor, a variant of decentralized SGD with differential privacy (DP) guarantees. Essentially, in Decor, users securely exchange randomness seeds in one communication round to generate pairwise-canceling correlated Gaussian noises, which are injected to protect local models at every communication round. We theoretically and empirically show that, for arbitrary connected graphs, Decor matches the central DP optimal privacy-utility trade-off. We do so under SecLDP, our new relaxation of local DP, which protects all user communications against an external eavesdropper and curious users, assuming that every pair of connected users shares a secret, i.e., an information hidden to all others. The main theoretical challenge is to control the accumulation of non-canceling correlated noise due to network sparsity. We also propose a companion SecLDP privacy accountant for public use., Comment: Accepted as conference paper at ICML 2024

The success of machine learning (ML) has been intimately linked with the availability of large amounts of data, typically collected from heterogeneous sources and processed on vast networks of computing devices (also called {\em workers}). Beyond accuracy, the use of ML in critical domains such as healthcare and autonomous driving calls for robustness against {\em data poisoning}and some {\em faulty workers}. The problem of {\em Byzantine ML} formalizes these robustness issues by considering a distributed ML environment in which workers (storing a portion of the global dataset) can deviate arbitrarily from the prescribed algorithm. Although the problem has attracted a lot of attention from a theoretical point of view, its practical importance for addressing realistic faults (where the behavior of any worker is locally constrained) remains unclear. It has been argued that the seemingly weaker threat model where only workers' local datasets get poisoned is more reasonable. We prove that, while tolerating a wider range of faulty behaviors, Byzantine ML yields solutions that are, in a precise sense, optimal even under the weaker data poisoning threat model. Then, we study a generic data poisoning model wherein some workers have {\em fully-poisonous local data}, i.e., their datasets are entirely corruptible, and the remainders have {\em partially-poisonous local data}, i.e., only a fraction of their local datasets is corruptible. We prove that Byzantine-robust schemes yield optimal solutions against both these forms of data poisoning, and that the former is more harmful when workers have {\em heterogeneous} local data., Comment: 38 pages

Byzantine agreement enables n processes to agree on a common L-bit value, despite up to t > 0 arbitrary failures. A long line of work has been dedicated to improving the bit complexity of Byzantine agreement in synchrony. This has culminated in COOL, an error-free (deterministically secure against a computationally unbounded adversary) solution that achieves O(nL + n^2 logn) worst-case bit complexity (which is optimal for L >= n logn according to the Dolev-Reischuk lower bound). COOL satisfies strong unanimity: if all correct processes propose the same value, only that value can be decided. Strong unanimity is, however, not sufficient for today's state machine replication (SMR) and blockchain protocols. These systems value progress and require a decided value to always be valid, excluding default decisions (such as EMPTY) even in cases where there is no unanimity a priori. Validated Byzantine agreement satisfies this property (called external validity). Yet, the best error-free (or even signature-free) validated agreement solutions achieve only O(n^2L) bit complexity, a far cry from the Omega(nL + n^2) Dolev-Reishcuk lower bound. In this paper, we present two new synchronous algorithms for validated Byzantine agreement, HashExt and ErrorFreeExt, with different trade-offs. Both algorithms are (1) signature-free, (2) optimally resilient (tolerate up to t < n / 3 failures), and (3) early-stopping (terminate in O(f+1) rounds, where f <= t is the actual number of failures). On the one hand, HashExt uses only hashes and achieves O(nL + n^3 kappa) bit complexity, which is optimal for L >= n^2 kappa (where kappa is the size of a hash). On the other hand, ErrorFreeExt is error-free, using no cryptography whatsoever, and achieves O( (nL + n^2) logn ) bit complexity, which is near-optimal for any L.

The possibility of adversarial (a.k.a., {\em Byzantine}) clients makes federated learning (FL) prone to arbitrary manipulation. The natural approach to robustify FL against adversarial clients is to replace the simple averaging operation at the server in the standard $\mathsf{FedAvg}$ algorithm by a \emph{robust averaging rule}. While a significant amount of work has been devoted to studying the convergence of federated {\em robust averaging} (which we denote by $\mathsf{FedRo}$), prior work has largely ignored the impact of {\em client subsampling} and {\em local steps}, two fundamental FL characteristics. While client subsampling increases the effective fraction of Byzantine clients, local steps increase the drift between the local updates computed by honest (i.e., non-Byzantine) clients. Consequently, a careless deployment of $\mathsf{FedRo}$ could yield poor performance. We validate this observation by presenting an in-depth analysis of $\mathsf{FedRo}$ tightly analyzing the impact of client subsampling and local steps. Specifically, we present a sufficient condition on client subsampling for nearly-optimal convergence of $\mathsf{FedRo}$ (for smooth non-convex loss). Also, we show that the rate of improvement in learning accuracy {\em diminishes} with respect to the number of clients subsampled, as soon as the sample size exceeds a threshold value. Interestingly, we also observe that under a careful choice of step-sizes, the learning error due to Byzantine clients decreases with the number of local steps. We validate our theory by experiments on the FEMNIST and CIFAR-$10$ image classification tasks.

Byzantine agreement allows n processes to decide on a common value, in spite of arbitrary failures. The seminal Dolev-Reischuk bound states that any deterministic solution to Byzantine agreement exchanges Omega(n^2) bits. In synchronous networks, solutions with optimal O(n^2) bit complexity, optimal fault tolerance, and no cryptography have been established for over three decades. However, these solutions lack robustness under adverse network conditions. Therefore, research has increasingly focused on Byzantine agreement for partially synchronous networks. Numerous solutions have been proposed for the partially synchronous setting. However, these solutions are notoriously hard to prove correct, and the most efficient cryptography-free algorithms still require O(n^3) exchanged bits in the worst case. In this paper, we introduce Oper, the first generic transformation of deterministic Byzantine agreement algorithms from synchrony to partial synchrony. Oper requires no cryptography, is optimally resilient (n >= 3t+1, where t is the maximum number of failures), and preserves the worst-case per-process bit complexity of the transformed synchronous algorithm. Leveraging Oper, we present the first partially synchronous Byzantine agreement algorithm that (1) achieves optimal O(n^2) bit complexity, (2) requires no cryptography, and (3) is optimally resilient (n >= 3t+1), thus showing that the Dolev-Reischuk bound is tight even in partial synchrony. Moreover, we adapt Oper for long values and obtain several new partially synchronous algorithms with improved complexity and weaker (or completely absent) cryptographic assumptions.

The success of machine learning (ML) applications relies on vast datasets and distributed architectures which, as they grow, present major challenges. In real-world scenarios, where data often contains sensitive information, issues like data poisoning and hardware failures are common. Ensuring privacy and robustness is vital for the broad adoption of ML in public life. This paper examines the costs associated with achieving these objectives in distributed ML architectures, from both theoretical and empirical perspectives. We overview the meanings of privacy and robustness in distributed ML, and clarify how they can be achieved efficiently in isolation. However, we contend that the integration of these two objectives entails a notable compromise in computational efficiency. In short, traditional noise injection hurts accuracy by concealing poisoned inputs, while cryptographic methods clash with poisoning defenses due to their non-linear nature. However, we outline future research directions aimed at reconciling this compromise with efficiency by considering weaker threat models.

Byzantine agreement, arguably the most fundamental problem in distributed computing, operates among n processes, out of which t < n can exhibit arbitrary failures. The problem states that all correct (non-faulty) processes must eventually decide (termination) the same value (agreement) from a set of admissible values defined by the proposals of the processes (validity). Depending on the exact version of the validity property, Byzantine agreement comes in different forms, from Byzantine broadcast to strong and weak consensus, to modern variants of the problem introduced in today's blockchain systems. Regardless of the specific flavor of the agreement problem, its communication cost is a fundamental metric whose improvement has been the focus of decades of research. The Dolev-Reischuk bound, one of the most celebrated results in distributed computing, proved 40 years ago that, at least for Byzantine broadcast, no deterministic solution can do better than Omega(t^2) exchanged messages in the worst case. Since then, it remained unknown whether the quadratic lower bound extends to seemingly weaker variants of Byzantine agreement. This paper answers the question in the affirmative, closing this long-standing open problem. Namely, we prove that any non-trivial agreement problem requires Omega(t^2) messages to be exchanged in the worst case. To prove the general lower bound, we determine the weakest Byzantine agreement problem and show, via a novel indistinguishability argument, that it incurs Omega(t^2) exchanged messages.

We present Epidemic Learning (EL), a simple yet powerful decentralized learning (DL) algorithm that leverages changing communication topologies to achieve faster model convergence compared to conventional DL approaches. At each round of EL, each node sends its model updates to a random sample of $s$ other nodes (in a system of $n$ nodes). We provide an extensive theoretical analysis of EL, demonstrating that its changing topology culminates in superior convergence properties compared to the state-of-the-art (static and dynamic) topologies. Considering smooth non-convex loss functions, the number of transient iterations for EL, i.e., the rounds required to achieve asymptotic linear speedup, is in $O(n^3/s^2)$ which outperforms the best-known bound $O(n^3)$ by a factor of $s^2$, indicating the benefit of randomized communication for DL. We empirically evaluate EL in a 96-node network and compare its performance with state-of-the-art DL approaches. Our results illustrate that EL converges up to $ 1.7\times$ quicker than baseline DL algorithms and attains $2.2 $\% higher accuracy for the same communication volume., Comment: Accepted paper at NeurIPS 2023

The theory underlying robust distributed learning algorithms, designed to resist adversarial machines, matches empirical observations when data is homogeneous. Under data heterogeneity however, which is the norm in practical scenarios, established lower bounds on the learning error are essentially vacuous and greatly mismatch empirical observations. This is because the heterogeneity model considered is too restrictive and does not cover basic learning tasks such as least-squares regression. We consider in this paper a more realistic heterogeneity model, namely (G,B)-gradient dissimilarity, and show that it covers a larger class of learning problems than existing theory. Notably, we show that the breakdown point under heterogeneity is lower than the classical fraction 1/2. We also prove a new lower bound on the learning error of any distributed learning algorithm. We derive a matching upper bound for a robust variant of distributed gradient descent, and empirically show that our analysis reduces the gap between theory and practice., Comment: Accepted to NeurIPS 2023

Due to the widespread availability of data, machine learning (ML) algorithms are increasingly being implemented in distributed topologies, wherein various nodes collaborate to train ML models via the coordination of a central server. However, distributed learning approaches face significant vulnerabilities, primarily stemming from two potential threats. Firstly, the presence of Byzantine nodes poses a risk of corrupting the learning process by transmitting inaccurate information to the server. Secondly, a curious server may compromise the privacy of individual nodes, sometimes reconstructing the entirety of the nodes' data. Homomorphic encryption (HE) has emerged as a leading security measure to preserve privacy in distributed learning under non-Byzantine scenarios. However, the extensive computational demands of HE, particularly for high-dimensional ML models, have deterred attempts to design purely homomorphic operators for non-linear robust aggregators. This paper introduces SABLE, the first homomorphic and Byzantine robust distributed learning algorithm. SABLE leverages HTS, a novel and efficient homomorphic operator implementing the prominent coordinate-wise trimmed mean robust aggregator. Designing HTS enables us to implement HMED, a novel homomorphic median aggregator. Extensive experiments on standard ML tasks demonstrate that SABLE achieves practical execution times while maintaining an ML accuracy comparable to its non-private counterpart.

The strong Byzantine agreement (SBA) problem is defined among n processes, out of which t < n can be faulty and behave arbitrarily. SBA allows correct (non-faulty) processes to agree on a common value. Moreover, if all correct processes have proposed the same value, only that value can be agreed upon. It has been known for a long time that any solution to the SBA problem incurs quadratic worst-case word complexity; additionally, the bound was known to be tight. However, no existing protocol achieves adaptive word complexity, where the number of exchanged words depends on the actual number of faults, and not on the upper bound. Therefore, it is still unknown whether SBA with adaptive word complexity exists. This paper answers the question in the affirmative. Namely, we introduce STRONG, a synchronous protocol that solves SBA among n = (2 + Omega(1))t + 1 processes and achieves adaptive word complexity. We show that the fundamental challenge of adaptive SBA lies in efficiently solving certification, the problem of obtaining a constant-sized, locally-verifiable proof that a value can safely be decided.

Detecting the source of a gossip is a critical issue, related to identifying patient zero in an epidemic, or the origin of a rumor in a social network. Although it is widely acknowledged that random and local gossip communications make source identification difficult, there exists no general quantification of the level of anonymity provided to the source. This paper presents a principled method based on $\varepsilon$-differential privacy to analyze the inherent source anonymity of gossiping for a large class of graphs. First, we quantify the fundamental limit of source anonymity any gossip protocol can guarantee in an arbitrary communication graph. In particular, our result indicates that when the graph has poor connectivity, no gossip protocol can guarantee any meaningful level of differential privacy. This prompted us to further analyze graphs with controlled connectivity. We prove on these graphs that a large class of gossip protocols, namely cobra walks, offers tangible differential privacy guarantees to the source. In doing so, we introduce an original proof technique based on the reduction of a gossip protocol to what we call a random walk with probabilistic die out. This proof technique is of independent interest to the gossip community and readily extends to other protocols inherited from the security community, such as the Dandelion protocol. Interestingly, our tight analysis precisely captures the trade-off between dissemination time of a gossip protocol and its source anonymity., Comment: Full version of DISC2023 paper

Consensus enables n processes to agree on a common valid L-bit value, despite t < n/3 processes being faulty and acting arbitrarily. A long line of work has been dedicated to improving the worst-case communication complexity of consensus in partial synchrony. This has recently culminated in the worst-case word complexity of O(n^2). However, the worst-case bit complexity of the best solution is still O(n^2 L + n^2 kappa) (where kappa is the security parameter), far from the \Omega(n L + n^2) lower bound. The gap is significant given the practical use of consensus primitives, where values typically consist of batches of large size (L > n). This paper shows how to narrow the aforementioned gap while achieving optimal linear latency. Namely, we present a new algorithm, DARE (Disperse, Agree, REtrieve), that improves upon the O(n^2 L) term via a novel dispersal primitive. DARE achieves O(n^{1.5} L + n^{2.5} kappa) bit complexity, an effective sqrt{n}-factor improvement over the state-of-the-art (when L > n kappa). Moreover, we show that employing heavier cryptographic primitives, namely STARK proofs, allows us to devise DARE-Stark, a version of DARE which achieves the near-optimal bit complexity of O(n L + n^2 poly(kappa)). Both DARE and DARE-Stark achieve optimal O(n) latency.

Whenever applicable, the Stochastic Gradient Descent (SGD) has shown itself to be unreasonably effective. Instead of underperforming and getting trapped in local minima due to the batch noise, SGD leverages it to learn to generalize better and find minima that are good enough for the entire dataset. This led to numerous theoretical and experimental investigations, especially in the context of Artificial Neural Networks (ANNs), leading to better machine learning algorithms. However, SGD is not applicable in a non-differentiable setting, leaving all that prior research off the table. In this paper, we show that a class of evolutionary algorithms (EAs) inspired by the Gillespie-Orr Mutational Landscapes model for natural evolution is formally equivalent to SGD in certain settings and, in practice, is well adapted to large ANNs. We refer to such EAs as Gillespie-Orr EA class (GO-EAs) and empirically show how an insight transfer from SGD can work for them. We then show that for ANNs trained to near-optimality or in the transfer learning setting, the equivalence also allows transferring the insights from the Mutational Landscapes model to SGD. We then leverage this equivalence to experimentally show how SGD and GO-EAs can provide mutual insight through examples of minima flatness, transfer learning, and mixing of individuals in EAs applied to large models., Comment: To be published in ALIFE 2023; 16 pages, 10 figures, 1 listing

Modern machine learning (ML) models are capable of impressive performances. However, their prowess is not due only to the improvements in their architecture and training algorithms but also to a drastic increase in computational power used to train them. Such a drastic increase led to a growing interest in distributed ML, which in turn made worker failures and adversarial attacks an increasingly pressing concern. While distributed byzantine resilient algorithms have been proposed in a differentiable setting, none exist in a gradient-free setting. The goal of this work is to address this shortcoming. For that, we introduce a more general definition of byzantine-resilience in ML - the \textit{model-consensus}, that extends the definition of the classical distributed consensus. We then leverage this definition to show that a general class of gradient-free ML algorithms - ($1,\lambda$)-Evolutionary Search - can be combined with classical distributed consensus algorithms to generate gradient-free byzantine-resilient distributed learning algorithms. We provide proofs and pseudo-code for two specific cases - the Total Order Broadcast and proof-of-work leader election., Comment: 10 pages, 4 listings, 2 theorems

The self-attention revolution allowed generative language models to scale and achieve increasingly impressive abilities. Such models - commonly referred to as Large Language Models (LLMs) - have recently gained prominence with the general public, thanks to conversational fine-tuning, putting their behavior in line with public expectations regarding AI. This prominence amplified prior concerns regarding the misuse of LLMs and led to the emergence of numerous tools to detect LLMs in the wild. Unfortunately, most such tools are critically flawed. While major publications in the LLM detectability field suggested that LLMs were easy to detect with fine-tuned autoencoders, the limitations of their results are easy to overlook. Specifically, they assumed publicly available generative models without fine-tunes or non-trivial prompts. While the importance of these assumptions has been demonstrated, until now, it remained unclear how well such detection could be countered. Here, we show that an attacker with access to such detectors' reference human texts and output not only evades detection but can fully frustrate the detector training - with a reasonable budget and all its outputs labeled as such. Achieving it required combining common "reinforcement from critic" loss function modification and AdamW optimizer, which led to surprisingly good fine-tuning generalization. Finally, we warn against the temptation to transpose the conclusions obtained in RNN-driven text GANs to LLMs due to their better representative ability. These results have critical implications for the detection and prevention of malicious use of generative language models, and we hope they will aid the designers of generative models and detectors., Comment: 15 pages, 6 figures; 10 pages, 7 figures Supplementary Materials; under review at ECML 2023

At the heart of state machine replication, the celebrated technique enabling decentralized and secure universal computation, lies Atomic Broadcast, a fundamental communication primitive that orders, authenticates, and deduplicates messages. This paper presents Chop Chop, a Byzantine Atomic Broadcast system that uses a novel authenticated memory pool to amortize the cost of ordering, authenticating and deduplicating messages, achieving "line rate" (i.e., closely matching the complexity of a protocol that does not ensure any ordering, authentication or Byzantine resilience) even when processing messages as small as 8 bytes. Chop Chop attains this performance by means of a new form of batching we call distillation. A distilled batch is a set of messages that are fast to authenticate, deduplicate, and order. Batches are distilled using a novel interactive protocol involving brokers, an untrusted layer of facilitating processes between clients and servers. In a geo-distributed deployment of 64 medium-sized servers, Chop Chop processes 43,600,000 messages per second with an average latency of 3.6 seconds. Under the same conditions, state-of-the-art alternatives offer two orders of magnitude less throughput for the same latency. We showcase three simple Chop Chop applications: a Payment system, an Auction house and a "Pixel war" game, respectively achieving 32, 2.3 and 35 million operations per second., Comment: Extended version of the paper appearing at OSDI 2024 with formal definitions, pseudocode, and proofs added in appendix

The ubiquity of distributed machine learning (ML) in sensitive public domain applications calls for algorithms that protect data privacy, while being robust to faults and adversarial behaviors. Although privacy and robustness have been extensively studied independently in distributed ML, their synthesis remains poorly understood. We present the first tight analysis of the error incurred by any algorithm ensuring robustness against a fraction of adversarial machines, as well as differential privacy (DP) for honest machines' data against any other curious entity. Our analysis exhibits a fundamental trade-off between privacy, robustness, and utility. To prove our lower bound, we consider the case of mean estimation, subject to distributed DP and robustness constraints, and devise reductions to centralized estimation of one-way marginals. We prove our matching upper bound by presenting a new distributed ML algorithm using a high-dimensional robust aggregation rule. The latter amortizes the dependence on the dimension in the error (caused by adversarial workers and DP), while being agnostic to the statistical properties of the data., Comment: Accepted paper at ICML

Byzantine machine learning (ML) aims to ensure the resilience of distributed learning algorithms to misbehaving (or Byzantine) machines. Although this problem received significant attention, prior works often assume the data held by the machines to be homogeneous, which is seldom true in practical settings. Data heterogeneity makes Byzantine ML considerably more challenging, since a Byzantine machine can hardly be distinguished from a non-Byzantine outlier. A few solutions have been proposed to tackle this issue, but these provide suboptimal probabilistic guarantees and fare poorly in practice. This paper closes the theoretical gap, achieving optimality and inducing good empirical results. In fact, we show how to automatically adapt existing solutions for (homogeneous) Byzantine ML to the heterogeneous setting through a powerful mechanism, we call nearest neighbor mixing (NNM), which boosts any standard robust distributed gradient descent variant to yield optimal Byzantine resilience under heterogeneity. We obtain similar guarantees (in expectation) by plugging NNM in the distributed stochastic heavy ball method, a practical substitute to distributed gradient descent. We obtain empirical results that significantly outperform state-of-the-art Byzantine ML solutions., Comment: Accepted paper at AISTATS 2023

The Byzantine consensus problem involves $n$ processes, out of which t < n could be faulty and behave arbitrarily. Three properties characterize consensus: (1) termination, requiring correct (non-faulty) processes to eventually reach a decision, (2) agreement, preventing them from deciding different values, and (3) validity, precluding ``unreasonable'' decisions. But, what is a reasonable decision? Strong validity, a classical property, stipulates that, if all correct processes propose the same value, only that value can be decided. Weak validity, another established property, stipulates that, if all processes are correct and they propose the same value, that value must be decided. The space of possible validity properties is vast. However, their impact on consensus remains unclear. This paper addresses the question of which validity properties allow Byzantine consensus to be solvable with partial synchrony, and at what cost. First, we determine necessary and sufficient conditions for a validity property to make the consensus problem solvable; we say that such validity properties are solvable. Notably, we prove that, if n <= 3t, all solvable validity properties are trivial (there exists an always-admissible decision). Furthermore, we show that, with any non-trivial (and solvable) validity property, consensus requires Omega(t^2) messages. This extends the seminal Dolev-Reischuk bound, originally proven for strong validity, to all non-trivial validity properties. Lastly, we give a general Byzantine consensus algorithm, we call Universal, for any solvable (and non-trivial) validity property. Importantly, Universal incurs O(n^2) message complexity. Thus, together with our lower bound, Universal implies a fundamental result in partial synchrony: with t \in Omega(n), the message complexity of all (non-trivial) consensus variants is Theta(n^2)., Comment: The extended version of the PODC 2023 paper

We propose uBFT, the first State Machine Replication (SMR) system to achieve microsecond-scale latency in data centers, while using only $2f{+}1$ replicas to tolerate $f$ Byzantine failures. The Byzantine Fault Tolerance (BFT) provided by uBFT is essential as pure crashes appear to be a mere illusion with real-life systems reportedly failing in many unexpected ways. uBFT relies on a small non-tailored trusted computing base -- disaggregated memory -- and consumes a practically bounded amount of memory. uBFT is based on a novel abstraction called Consistent Tail Broadcast, which we use to prevent equivocation while bounding memory. We implement uBFT using RDMA-based disaggregated memory and obtain an end-to-end latency of as little as 10us. This is at least 50$\times$ faster than MinBFT , a state of the art $2f{+}1$ BFT SMR based on Intel's SGX. We use uBFT to replicate two KV-stores (Memcached and Redis), as well as a financial order matching engine (Liquibook). These applications have low latency (up to 20us) and become Byzantine tolerant with as little as 10us more. The price for uBFT is a small amount of reliable disaggregated memory (less than 1 MiB), which in our prototype consists of a small number of memory servers connected through RDMA and replicated for fault tolerance.

Blockchain intercommunication systems enable the exchanges of messages between blockchains. This interoperability promotes innovation, unlocks liquidity and access to assets. However, blockchains are isolated systems that originally were not designed for interoperability. This makes cross-chain communication, or bridges for short, insecure by nature. More precisely, cross-chain systems face security challenges in terms of selfish rational players such as maximal extractable value (MEV) and censorship. We propose to solve these challenges using zero knowledge proofs (ZKPs) for cross-chain communication. Securing cross-chain communication is remarkably more complex than securing single-chain events as such a system must preserve user security against both on- and off-chain analysis. To achieve this goal, we propose the following pair of contributions: the DACT protocol and the SurferMonkey infrastructure that supports the DACT protocol. The DACT protocol is a global solution for the anonymity and security challenges of agnostic blockchain intercommunication. DACT breaks on- and off-chain analysis thanks to the use of ZKPs. SurferMonkey is a decentralized infrastructure that implements DACT in practice. Since SurferMonkey works at the blockchain application layer, any decentralized application (dApp) can use SurferMonkey to send any type of message to a dApp on another blockchain. With SurferMonkey, users can neither be censored nor be exposed to MEV. By applying decentralized proactive security, we obtain resilience against selfish rational players, and raise the security bar against cyberattacks. We have implemented a proof of concept (PoC) of SurferMonkey by reverse engineering Tornado Cash and by applying IDEN3 ZKP circuits. SurferMonkey enables new usecases, ranging from anonymous voting and gaming, to a new phase of anonymous decentralized finance (aDeFi).

Storage disaggregation underlies today's cloud and is naturally complemented by pushing down some computation to storage, thus mitigating the potential network bottleneck between the storage and compute tiers. We show how ML training benefits from storage pushdowns by focusing on transfer learning (TL), the widespread technique that democratizes ML by reusing existing knowledge on related tasks. We propose HAPI, a new TL processing system centered around two complementary techniques that address challenges introduced by disaggregation. First, applications must carefully balance execution across tiers for performance. HAPI judiciously splits the TL computation during the feature extraction phase yielding pushdowns that not only improve network time but also improve total TL training time by overlapping the execution of consecutive training iterations across tiers. Second, operators want resource efficiency from the storage-side computational resources. HAPI employs storage-side batch size adaptation allowing increased storage-side pushdown concurrency without affecting training accuracy. HAPI yields up to 2.5x training speed-up while choosing in 86.8% of cases the best performing split point or one that is at most 5% off from the best., Comment: To appear in the proceedings of SoCC '24

Large AI Models (LAIMs), of which large language models are the most prominent recent example, showcase some impressive performance. However they have been empirically found to pose serious security issues. This paper systematizes our knowledge about the fundamental impossibility of building arbitrarily accurate and secure machine learning models. More precisely, we identify key challenging features of many of today's machine learning settings. Namely, high accuracy seems to require memorizing large training datasets, which are often user-generated and highly heterogeneous, with both sensitive information and fake users. We then survey statistical lower bounds that, we argue, constitute a compelling case against the possibility of designing high-accuracy LAIMs with strong security guarantees., Comment: 40 pages

Byzantine Reliable Broadcast (BRB) is a fundamental distributed computing primitive, with applications ranging from notifications to asynchronous payment systems. Motivated by practical consideration, we study Client-Server Byzantine Reliable Broadcast (CSB), a multi-shot variant of BRB whose interface is split between broadcasting clients and delivering servers. We present Draft, an optimally resilient implementation of CSB. Like most implementations of BRB, Draft guarantees both liveness and safety in an asynchronous environment. Under good conditions, however, Draft achieves unparalleled efficiency. In a moment of synchrony, free from Byzantine misbehaviour, and at the limit of infinitely many broadcasting clients, a Draft server delivers a $b$-bits payload at an asymptotic amortized cost of $0$ signature verifications, and $\log_2(c) + b$ bits exchanged, where $c$ is the number of clients in the system. This is the information-theoretical minimum number of bits required to convey the payload ($b$ bits, assuming it is compressed), along with an identifier for its sender ($\log_2(c)$ bits, necessary to enumerate any set of $c$ elements, and optimal if broadcasting frequencies are uniform or unknown). These two achievements have profound practical implications. Real-world BRB implementations are often bottlenecked either by expensive signature verifications, or by communication overhead. For Draft, instead, the network is the limit: a server can deliver payloads as quickly as it would receive them from an infallible oracle.

Collaborative learning algorithms, such as distributed SGD (or D-SGD), are prone to faulty machines that may deviate from their prescribed algorithm because of software or hardware bugs, poisoned data or malicious behaviors. While many solutions have been proposed to enhance the robustness of D-SGD to such machines, previous works either resort to strong assumptions (trusted server, homogeneous data, specific noise model) or impose a gradient computational cost that is several orders of magnitude higher than that of D-SGD. We present MoNNA, a new algorithm that (a) is provably robust under standard assumptions and (b) has a gradient computation overhead that is linear in the fraction of faulty machines, which is conjectured to be tight. Essentially, MoNNA uses Polyak's momentum of local gradients for local updates and nearest-neighbor averaging (NNA) for global mixing, respectively. While MoNNA is rather simple to implement, its analysis has been more challenging and relies on two key elements that may be of independent interest. Specifically, we introduce the mixing criterion of $(\alpha, \lambda)$-reduction to analyze the non-linear mixing of non-faulty machines, and present a way to control the tension between the momentum and the model drifts. We validate our theory by experiments on image classification and make our code available at https://github.com/LPD-EPFL/robust-collaborative-learning., Comment: Accepted paper at ICML 2023

This paper introduces Carbon, a high-throughput system enabling asynchronous (safe) and consensus-free (efficient) payments and votes within a dynamic set of clients. Carbon is operated by a dynamic set of validators that may be reconfigured asynchronously, offering its clients eclipse resistance as well as lightweight bootstrap. Carbon offers clients the ability to select validators by voting them in and out of the system thanks to its novel asynchronous and stake-less voting mechanism. Carbon relies on an asynchronous and deterministic implementation of Byzantine reliable broadcast that uniquely leverages a permissionless set of untrusted servers, brokers, to slash the cost of client authentication inherent to Byzantine fault tolerant systems. Carbon is able to sustain a throughput of one million payments per second in a geo-distributed environment, outperforming the state of the art by three orders of magnitude with equivalent latencies., Comment: This is an extended version of the paper appearing at IEEE TDSC 2024 under DOI 10.1109/TDSC.2024.3428617 with formal definitions, pseudocode, and proofs added in appendices; these appendices correspond to the previous version of this paper on arXiv (arXiv:2209.09580v2)

The Dolev-Reischuk bound says that any deterministic Byzantine consensus protocol has (at least) quadratic communication complexity in the worst case. While it has been shown that the bound is tight in synchronous environments, it is still unknown whether a consensus protocol with quadratic communication complexity can be obtained in partial synchrony. Until now, the most efficient known solutions for Byzantine consensus in partially synchronous settings had cubic communication complexity (e.g., HotStuff, binary DBFT). This paper closes the existing gap by introducing SQuad, a partially synchronous Byzantine consensus protocol with quadratic worst-case communication complexity. In addition, SQuad is optimally-resilient and achieves linear worst-case latency complexity. The key technical contribution underlying SQuad lies in the way we solve view synchronization, the problem of bringing all correct processes to the same view with a correct leader for sufficiently long. Concretely, we present RareSync, a view synchronization protocol with quadratic communication complexity and linear latency complexity, which we utilize in order to obtain SQuad.