Secure Coded Distributed Computing and Extensions to Multiple Access Setting
Shanuja Sasi, Onur Günlü · IEEE Transactions on Communications · 2025
We consider two critical aspects of security in thedistributed computing (DC)model:secure data shufflingandsecure coded computing. It is imperative that any external entity overhearing the communication does not gain any information about theintermediate values (IVs)exchanged during the shuffling phase of the DC model. Our approach ensures IV confidentiality during data shuffling. Moreover, each node in the system must be able to recover the IVs necessary for computing its output functions but must also remain oblivious to the IVs associated with output functions not assigned to it. We design secure DC methods and establish achievable limits on the tradeoffs between the communication and computation loads to contribute to the advancement of secure data processing in distributed systems. First, we establish that the computation and communication loads stay the same as for non-secure data shuffling. However, implementing secure data shuffling requires additional overhead for storing secret keys at the nodes. Next, we show that for secure coded computation, both the computation and communication loads increase compared to the non-secure scenario, along with the overhead for storing secret keys. Finally, we extend our security results to a novel distributed computing model known asmulti-access distributed computing (MADC), which was recently introduced. The MADC model features two distinct sets of nodes, namelymapperandreducernodes. Unlike the original setting where mapper and reducer nodes were the same, in this model, they are separate entities, and each reducer node is connected to multiple mapper nodes. We show that, for MADC models also, computation and communication loads remain the same with or without secure data shuffling. However, secure coded computation results in increased computation and communication loads compared to the non-secure case, and both scenarios require overhead for storing secret keys at the reducer nodes.