Diksha Gupta

Diksha Gupta, Carlos D. Brody

Trial history biases in decision-making tasks are thought to reflect systematic updates of decision variables, therefore their precise nature informs conclusions about underlying heuristic strategies and learning processes. However, random drifts in decision variables can corrupt this inference by mimicking the signatures of systematic updates. Hence, identifying the trial-by-trial evolution of decision variables requires methods that can robustly account for such drifts. Recent studies (Lak'20, Mendonça'20) have made important advances in this direction, by proposing a convenient method to correct for the influence of slow drifts in decision criterion, a key decision variable. Here we apply this correction to a variety of updating scenarios, and evaluate its performance. We show that the correction fails for a wide range of commonly assumed systematic updating strategies, distorting one's inference away from the veridical strategies towards a narrow subset. To address these limitations, we propose a model-based approach for disambiguating systematic updates from random drifts, and demonstrate its success on real and synthetic datasets. We show that this approach accurately recovers the latent trajectory of drifts in decision criterion as well as the generative systematic updates from simulated data. Our results offer recommendations for methods to account for the interactions between history biases and slow drifts, and highlight the advantages of incorporating assumptions about the generative process directly into models of decision-making.

Diksha Gupta, Antonio Honsell, Chuan Xu et al.

Distributed learning enables scalable model training over decentralized data, but remains hindered by Byzantine faults and high communication costs. While both challenges have been studied extensively in isolation, their interplay has received limited attention. Prior work has shown that naively combining communication compression with Byzantine-robust aggregation can severely weaken resilience to faulty nodes. The current state-of-the-art, Byz-DASHA-PAGE, leverages a momentum-based variance reduction scheme to counteract the negative effect of compression noise on Byzantine robustness. In this work, we introduce RoSDHB, a new algorithm that integrates classical Polyak momentum with a coordinated compression strategy. Theoretically, RoSDHB matches the convergence guarantees of Byz-DASHA-PAGE under the standard $(G,B)$-gradient dissimilarity model, while relying on milder assumptions and requiring less memory and communication per client. Empirically, RoSDHB demonstrates stronger robustness while achieving substantial communication savings compared to Byz-DASHA-PAGE.

Diksha Gupta, Jared Saia, Maxwell Young

A Sybil attack occurs when an adversary controls multiple identifiers (IDs) in a system. Limiting the number of Sybil (bad) IDs to a minority is critical to the use of well-established tools for tolerating malicious behavior, such as Byzantine agreement and secure multiparty computation. A popular technique for enforcing a Sybil minority is resource burning: the verifiable consumption of a network resource, such as computational power, bandwidth, or memory. Unfortunately, typical defenses based on resource burning require non-Sybil (good) IDs to consume at least as many resources as the adversary. Additionally, they have a high resource burning cost, even when the system membership is relatively stable. Here, we present a new Sybil defense, ERGO, that guarantees (1) there is always a minority of bad IDs; and (2) when the system is under significant attack, the good IDs consume asymptotically less resources than the bad. In particular, for churn rate that can vary exponentially, the resource burning rate for good IDs under ERGO is O(\sqrt{TJ} + J), where T is the resource burning rate of the adversary, and J is the join rate of good IDs. We show this resource burning rate is asymptotically optimal for a large class of algorithms. We empirically evaluate ERGO alongside prior Sybil defenses. Additionally, we show that ERGO can be combined with machine learning techniques for classifying Sybil IDs, while preserving its theoretical guarantees. Based on our experiments comparing ERGO with two previous Sybil defenses, ERGO improves on the amount of resource burning relative to the adversary by up to 2 orders of magnitude without machine learning, and up to 3 orders of magnitude using machine learning.

Diksha Gupta, Jared Saia, Maxwell Young

Proof-of-work (PoW) is an algorithmic tool used to secure networks by imposing a computational cost on participating devices. Unfortunately, traditional PoW schemes require that correct devices perform computational work perpetually, even when the system is not under attack. We address this issue by designing a general PoW protocol that ensures two properties. First, the network stays secure. In particular, the fraction of identities in the system that are controlled by an attacker is always less than 1/2. Second, our protocol's computational cost is commensurate with the cost of an attacker. In particular, the total computational cost of correct devices is a linear function of the attacker's computational cost plus the number of correct devices that have joined the system. Consequently, if the network is attacked, we ensure security with cost that grows linearly with the attacker's cost; and, in the absence of attack, our computational cost remains small. We prove similar guarantees for bandwidth cost. Our results hold in a dynamic, decentralized system where participants join and depart over time, and where the total computational power of the attacker is up to a constant fraction of the total computational power of correct devices. We demonstrate how to leverage our results to address important security problems in distributed computing including: Sybil attacks, Byzantine consensus, and Committee election.