X-Secure T-Private Federated Submodel Learning With Elastic Dropout Resilience

Motivated by recent interest in federated submodel learning, this work explores the fundamental problem of privately reading from and writing to a database comprised of <inline-formula> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> files (submodels) that are stored across <inline-formula> <tex-math notation="LaTeX">$N$ </tex-math></inline-formula> distributed servers according to an <inline-formula> <tex-math notation="LaTeX">$X$ </tex-math></inline-formula>-secure threshold secret sharing scheme. One after another, various users wish to retrieve their desired file, locally process the information and then update the file in the distributed database while keeping the identity of their desired file private from any set of up to <inline-formula> <tex-math notation="LaTeX">$T$ </tex-math></inline-formula> colluding servers. The availability of servers changes over time, so elastic dropout resilience is required. The main contribution of this work is an adaptive scheme, called ACSA-RW, that takes advantage of all currently available servers to reduce its communication costs, fully updates the database after each write operation even though the database is only partially accessible due to server dropouts, and ensures a memoryless operation of the network in the sense that the storage structure is preserved and future users may remain oblivious of the past history of server dropouts. The ACSA-RW construction builds upon cross-subspace alignment (CSA) codes that were originally introduced for <inline-formula> <tex-math notation="LaTeX">$X$ </tex-math></inline-formula>-secure <inline-formula> <tex-math notation="LaTeX">$T$ </tex-math></inline-formula>-private information retrieval and have been shown to be natural solutions for secure distributed batch matrix multiplication problems. ACSA-RW achieves the desired private read and write functionality with elastic dropout resilience, matches the best results for private-read from PIR literature, improves significantly upon available baselines for private-write, reveals a striking symmetry between upload and download costs, and exploits storage redundancy to accommodate arbitrary read and write dropout servers up to certain threshold values. It also answers in the affirmative an open question by Kairouz et al. for the case of partially colluding servers (i.e., tolerating collusion up to a threshold) by exploiting synergistic gains from the joint design of private read and write operations.