Now 6205 visitors
Today:0 Yesterday:0
Total: 152383 212S 50P 4R
2026-10-20, Week 43
Member Login
TACT Journal Homepage
Call for Paper
Committee
Find My Paper
Author Homepage
Statistics & History
Welcome Message
Paper Procedure
FAQ
Paper Archives
Outstanding Papers
Program & Proceedings
Presentation Platform
Hotel & Travel Info
Photo Gallery
Scheduler Login
Seminar
Archives Login
Sponsors


















Virual Oral Question and Answer
IEEE/ICACT20260329 Question.11
Questioner: shil@dsp.ac.cn    2026-02-09 오전 11:57:59 		IEEE/ICACT20260329 Answer.11
Answer by Auhor yangyn@dsp.ac.cn   2026-02-09 오전 11:57:59
How does CDynP handle privacy concerns while coordinating distributed training across multiple data centers, and what specific information is shared between data centers in the compressed state summary (SS)? CDynP addresses privacy concerns through a decentralized, chain-based coordination mechanism that avoids exposing any internal data center information. The specific information shared between DCs is the compressed state summary (SS), which is mathematically defined as SSk−1​[s][l]=min1≤g≤Gk−1​​dpk−1​[s][l][g]. This SS contains only the minimal maximum stage execution timefor partitioning a given number of layers (l) into a given number of stages (s). It is a highly aggregated cost metric that deliberately excludes all sensitive internal details, such as: The specific GPU allocation (g) that achieved that minimal time. Intra-DC topology or bandwidth (Bkintra​). Detailed workload distribution. During the "Forward CDynP" phase, this SS is passed sequentially from one DC to the next. Each DC uses the received SS to initialize its local dynamic programming process, makes a local optimization decision for its segment of the model, and then compresses its results into a new SS for the next hop. This process ensures that no DC ever has visibility into anothers internal resources or configuration, achieving privacy-preserving coordination. The theoretical analysis in the paper proves that this method can achieve a solution quality equivalent to a full-information approach under the privacy constraints.

Select Voice