Implement secure horizontal scheme for federated learning #10231
Conversation
…ute under secure scenario
…valent to broadcast
…lobal best split, but need to further apply split correctly
…w histogram transmission data structure of a flat vector
src/processing/processor.h
Outdated
|
|
||
| Processor* load(const std::string& plugin_name); | ||
|
|
||
| void unload(); |
There was a problem hiding this comment.
I couldn't find any reference to this unload method? Is it leaking the handle?
| } | ||
|
|
||
| // Save pairs for future operations | ||
| this->gh_pairs_ = new std::vector<double>(pairs); |
src/tree/hist/evaluate_splits.h
Outdated
| @@ -429,7 +430,7 @@ class HistEvaluator { | |||
| all_entries[worker * entries.size() + nidx_in_set].split); | |||
| } | |||
| } | |||
| if (is_secure_) { | |||
| if (is_secure_ && is_col_split_) { | |||
There was a problem hiding this comment.
Looking at the code, the is_col_split_ must be true.
There was a problem hiding this comment.
ah indeed, thanks!
removed the redundant condition.
Support bool params in proc_params.
| @@ -180,47 +221,100 @@ class HistogramBuilder { | |||
|
|
|||
| common::BlockedSpace2d space( | |||
| nodes_to_build.size(), [&](std::size_t) { return n_total_bins; }, 1024); | |||
| common::ParallelFor2d(space, this->n_threads_, [&](size_t node, common::Range1d r) { | |||
| common::ParallelFor2d(space, this->n_threads_, [&](std::size_t node, common::Range1d r) { | |||
| // Merging histograms from each thread. | |||
| this->buffer_.ReduceHist(node, r.begin(), r.end()); | |||
There was a problem hiding this comment.
Is this reduce hist function call actually useful for secure training?
There was a problem hiding this comment.
I think so, this is the multi-thread part for each client
There was a problem hiding this comment.
I thought the histogram is being built in the plugin? If so, any operation for the histogram in XGBoost should be invalid right?
There was a problem hiding this comment.
ah you are right!
this reduce hist will have no impact for secure vertical training, it will be needed by secure horizontal, and non-secure federated schemes.
trivialfis
merged commit de4013f
into
dmlc:vertical-federated-learning
Implementation for the secure horizontal scheme (RFC #10170).
Modifictions added beyond #10124
Following the same processor interface design with new functions to perform secure histogram aggregation (at external server).