Sandeep S. Kulkarni
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Sandeep S. Kulkarni has a total of 73 co-authors. Here are the top co-authors by number of shared publications:
- Borzoo Bonakdarpour
University of Waterloo
9 shared publications - Duong Nguyen
Michigan State University University at Buffalo SUNY Michigan State University University at Buffalo SUNY
9 shared publications - Murat Demirbas
Computer Science Department Computer Science Department University at Buffalo SUNY University at Buffalo SUNY
9 shared publications - Arya Tanmay Gupta
Michigan State University, East Lansing, United States
8 shared publications - Fuad Abujarad
Michigan State University
6 shared publications - Reza Hajisheykhi
Michigan State University
6 shared publications
Latest Publications
- Open Access
- 18 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 62-72
Simulating Delay Tolerant Networking with ns-3
- 18 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 62-72
Open AccessSimulating Delay Tolerant Networking with ns-3
- 15 August 2026
- Published by Springer Nature
- p. 556-570
Asynchronous Checkpoint for Eventually Consistent Databases
- 15 August 2026
- Published by Springer Nature
- p. 556-570
Asynchronous Checkpoint for Eventually Consistent Databases
- Open Access
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Replay Clocks for Resource Constrained Distributed Systems
Distributed systems generate partial orders of events, where only some relationships, e.g., send and receive of a message, are ordered, while many events remain unordered. External observers, however, perceive executions as a total order that extends this partial order. This discrepancy allows executions that appear correct to mask underlying violations that arise in unobserved interleavings. The challenge becomes even more complex in systems that rely on partially synchronized clocks and timing-based mechanisms such as leases, timeouts, and token expiration, where event ordering depends on both communication and clock skew assumptions. Achieving high‐fidelity replays under these conditions is difficult, particularly in resource‐constrained environments that cannot support heavy logging. This paper introduces replay clocks (RepCl), a timestamping mechanism designed for constrained distributed systems such as LoRaWAN. Unlike existing approaches that either produce impossible replays or fail to reproduce valid ones, RepCl captures the essential ordering information needed to generate all feasible replays while avoiding spurious ones. We show that RepCl significantly reduces timestamp space overhead with only a modest increase in computation cost. The resulting savings in storage and transmission energy substantially outweigh the additional computation, making RepCl a practical and efficient foundation for replay generation in resource constrained distributed systems.
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessReplay Clocks for Resource Constrained Distributed Systems
Distributed systems generate partial orders of events, where only some relationships, e.g., send and receive of a message, are ordered, while many events remain unordered. External observers, however, perceive executions as a total order that extends this partial order. This discrepancy allows executions that appear correct to mask underlying violations that arise in unobserved interleavings. The challenge becomes even more complex in systems that rely on partially synchronized clocks and timing-based mechanisms such as leases, timeouts, and token expiration, where event ordering depends on both communication and clock skew assumptions. Achieving high‐fidelity replays under these conditions is difficult, particularly in resource‐constrained environments that cannot support heavy logging. This paper introduces replay clocks (RepCl), a timestamping mechanism designed for constrained distributed systems such as LoRaWAN. Unlike existing approaches that either produce impossible replays or fail to reproduce valid ones, RepCl captures the essential ordering information needed to generate all feasible replays while avoiding spurious ones. We show that RepCl significantly reduces timestamp space overhead with only a modest increase in computation cost. The resulting savings in storage and transmission energy substantially outweigh the additional computation, making RepCl a practical and efficient foundation for replay generation in resource constrained distributed systems.
- Open Access
- 31 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Don't Trust the Hidden Gradients!
In this paper we propose novel methods for the computation of several statistical functions including mean, variance, and selective mean (over values that fit a specified criteria) that can be used to compute the outlier-resistant mean. We use these new tools to mitigate poisoning attacks in privacy-preserving cross-silo federated learning systems. Our methods are fully privacy-preserving such that no process learns the initial value of any other process during the protocol. Our approach is based on FHE and supports SIMD operations, making it scale efficiently with large amounts of data. Our approach allows clients to drop in/out during the protocol which is not supported by existing methods, such as garbled circuits. Our protocols support a trade-off between degree of privacy and efficiency, meaning that in cases where weaker privacy is required, more efficient versions of our protocols may be employed. We augment our protocols with collusion-resistance. Given two honest processes, our protocols can tolerate any number of colluding processes. Our protocols are evaluated in simulated cross-silo federated learning experiments. To our knowledge, this is the first privacy-preserving selective mean protocol.
- 31 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessDon't Trust the Hidden Gradients!
In this paper we propose novel methods for the computation of several statistical functions including mean, variance, and selective mean (over values that fit a specified criteria) that can be used to compute the outlier-resistant mean. We use these new tools to mitigate poisoning attacks in privacy-preserving cross-silo federated learning systems. Our methods are fully privacy-preserving such that no process learns the initial value of any other process during the protocol. Our approach is based on FHE and supports SIMD operations, making it scale efficiently with large amounts of data. Our approach allows clients to drop in/out during the protocol which is not supported by existing methods, such as garbled circuits. Our protocols support a trade-off between degree of privacy and efficiency, meaning that in cases where weaker privacy is required, more efficient versions of our protocols may be employed. We augment our protocols with collusion-resistance. Given two honest processes, our protocols can tolerate any number of colluding processes. Our protocols are evaluated in simulated cross-silo federated learning experiments. To our knowledge, this is the first privacy-preserving selective mean protocol.
- Open Access
- 18 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 70-81
Clock Skew Models for ns-3
- 18 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 70-81
Open AccessClock Skew Models for ns-3
- Open Access
SONNI: Secure Oblivious Neural Network Inference
In the standard privacy-preserving Machine learning as-a-service (MLaaS) model, the client encrypts data using homomorphic encryption and uploads it to a server for computation. The result is then sent back to the client for decryption. It has become more and more common for the computation to be outsourced to third-party servers. In this paper we identify a weakness in this protocol that enables a completely undetectable novel model-stealing attack that we call the Silver Platter attack. This attack works even under multikey encryption that prevents a simple collusion attack to steal model parameters. We also propose a mitigation that protects privacy even in the presence of a malicious server and malicious client or model provider (majority dishonest). When compared to a state-of-the-art but small encrypted model with 32k parameters, we preserve privacy with a failure chance of 1.51 x 10^-28 while batching capability is reduced by 0.2%. Our approach uses a novel results-checking protocol that ensures the computation was performed correctly without violating honest clients' data privacy. Even with collusion between the client and the server, they are unable to steal model parameters. Additionally, the model provider cannot learn any client data if maliciously working with the server.
Open AccessSONNI: Secure Oblivious Neural Network Inference
In the standard privacy-preserving Machine learning as-a-service (MLaaS) model, the client encrypts data using homomorphic encryption and uploads it to a server for computation. The result is then sent back to the client for decryption. It has become more and more common for the computation to be outsourced to third-party servers. In this paper we identify a weakness in this protocol that enables a completely undetectable novel model-stealing attack that we call the Silver Platter attack. This attack works even under multikey encryption that prevents a simple collusion attack to steal model parameters. We also propose a mitigation that protects privacy even in the presence of a malicious server and malicious client or model provider (majority dishonest). When compared to a state-of-the-art but small encrypted model with 32k parameters, we preserve privacy with a failure chance of 1.51 x 10^-28 while batching capability is reduced by 0.2%. Our approach uses a novel results-checking protocol that ensures the computation was performed correctly without violating honest clients' data privacy. Even with collusion between the client and the server, they are unable to steal model parameters. Additionally, the model provider cannot learn any client data if maliciously working with the server.
- Open Access
- 4 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 274-275
Quantitative characterization of consistency violating faults at individual distributed computing processes
- 4 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 274-275
Open AccessQuantitative characterization of consistency violating faults at individual distributed computing processes
- Open Access
- 4 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 280-281
Distributing Context-Aware Shared Memory Data Structures: A Case Study on Singly-Linked Lists
- 4 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 280-281
Open AccessDistributing Context-Aware Shared Memory Data Structures: A Case Study on Singly-Linked Lists
- 18 October 2024
- Published by Elsevier in Theoretical Computer Science
- Vol. 1024
Tolerance to asynchrony in algorithms for multiplication and modulo
- 18 October 2024
- Published by Elsevier in Theoretical Computer Science
- Vol. 1024
Tolerance to asynchrony in algorithms for multiplication and modulo
- Open Access
Tolerance to Asynchrony of an Algorithm for Gathering Myopic Robots on an Infinite Triangular Grid
In this paper, we study the problem of gathering distance-1 myopic robots on an infinite triangular grid. We show that the algorithm developed by Goswami et al. (SSS, 2022) is lattice-linear (cf. Gupta and Kulkarni, SRDS 2023). This implies that a distributed scheduler, assumed therein, is not required for this algorithm: it runs correctly in asynchrony. It also implies that the algorithm works correctly even if the robots are equipped with a unidirectional \textit{camera} to see the neighbouring robots (rather than an omnidirectional one, which would be required under a distributed scheduler). Due to lattice-linearity, we can predetermine the point of gathering. We also show that this algorithm converges in $2n$ rounds, which is lower than the complexity ($2.5(n+1)$ rounds) that was shown in Goswami et al.Open AccessTolerance to Asynchrony of an Algorithm for Gathering Myopic Robots on an Infinite Triangular Grid
In this paper, we study the problem of gathering distance-1 myopic robots on an infinite triangular grid. We show that the algorithm developed by Goswami et al. (SSS, 2022) is lattice-linear (cf. Gupta and Kulkarni, SRDS 2023). This implies that a distributed scheduler, assumed therein, is not required for this algorithm: it runs correctly in asynchrony. It also implies that the algorithm works correctly even if the robots are equipped with a unidirectional \textit{camera} to see the neighbouring robots (rather than an omnidirectional one, which would be required under a distributed scheduler). Due to lattice-linearity, we can predetermine the point of gathering. We also show that this algorithm converges in $2n$ rounds, which is lower than the complexity ($2.5(n+1)$ rounds) that was shown in Goswami et al.