PrivPSV

PrivPSV is a privacy-preserving party selection and valuation framework for vertical federated learning (VFL). It lets one active party select passive parties with informative features and evaluate their contributions with secure multi-party computation.

The implementation uses replicated secret sharing through MP-SPDZ. The C++ client program prepares local inputs and sends shares to MP-SPDZ, while the MP-SPDZ programs run the secure mutual information, feature selection, and valuation logic.

Prerequisites

• Ubuntu 20.04 or a compatible Linux environment
• MP-SPDZ under third_party/MP-SPDZ
• Boost
• CMake and a C++ compiler
• Docker, only if running in Docker mode

Check CMakeLists.txt before building and make sure SPDZ_HOME points to your local MP-SPDZ directory.

Build

From the repository root:

bash build.sh

This creates the C++ client executable at:

build/ppdv

Compile The Main MP-SPDZ Program

The main secure feature selection and valuation program is:

third_party/MP-SPDZ/Programs/Source/client_input_feature_selection.mpc

Compile it from the MP-SPDZ directory:

cd third_party/MP-SPDZ
./compile.py -R 64 -l Programs/Source/client_input_feature_selection.mpc -- \
  <n_rounds> <sample_num> <active_feature_num> <passive_feature_num> \
  <active_nbits> <passive_nbits> <threads> <class_num> <method_id> <k> \
  <inv_n> <n1> <n2> <n3>

Example for the current 1-active-party/3-passive-party bank setup:

cd third_party/MP-SPDZ
./compile.py -R 64 -l Programs/Source/client_input_feature_selection.mpc -- \
  1 2115 3 13 3 3 16 2 4 2 0.00009454 5 4 4

The generated program name is based on the file name and compile arguments, for example:

client_input_feature_selection-1-2115-3-13-3-3-16-2-4-2-0.00009454-5-4-4

Main Program Arguments

Argument	Meaning
n_rounds	Number of protocol rounds, usually 1.
sample_num	Number of samples held by each party.
active_feature_num	Number of active-party features.
passive_feature_num	Total passive-party features.
active_nbits	Bit length for active-party discretized values.
passive_nbits	Bit length for passive-party discretized values.
threads	Number of MP-SPDZ threads.
class_num	Number of label classes.
method_id	Secure computation method.
k	Number of top features selected by each passive party.
inv_n	Inverse sample count, 1 / sample_num.
n1, n2, n3	Feature count for passive parties.

Supported method_id values:

Method	Description
1	Sorting-based mutual information, treating all active features as one variable.
2	Indicator-vector-based mRMR feature selection.
3	Sorting-based mRMR feature selection.
4	Sorting-based feature selection and party valuation.

Run PrivPSV

Start the three MP-SPDZ parties first. Use the generated program name from the compile step.

For local execution:

cd third_party/MP-SPDZ
./replicated-ring-party.x -I -p 0 client_input_feature_selection-1-2115-3-13-3-3-16-2-4-2-0.00009454-5-4-4
./replicated-ring-party.x -I -p 1 client_input_feature_selection-1-2115-3-13-3-3-16-2-4-2-0.00009454-5-4-4
./replicated-ring-party.x -I -p 2 client_input_feature_selection-1-2115-3-13-3-3-16-2-4-2-0.00009454-5-4-4

For remote execution, add the MP-SPDZ host:

./replicated-ring-party.x -I -h <spdz-host> -p <0|1|2> <program-name>

Then run the active party and three passive parties from build/, one terminal per party:

cd build

./ppdv --party-id 0 --party-num 4 --party-type 0 \
  --network-file ../data/network/Parties-4.txt \
  --log-file ../log/ \
  --data-input-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/active_party.csv \
  --data-output-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/client0_output.txt \
  --mpc-party-num 3 --sample-num 2115 --class-num 2 \
  --method-id 4 --selected-feature-num 2

./ppdv --party-id 1 --party-num 4 --party-type 1 \
  --network-file ../data/network/Parties-4.txt \
  --log-file ../log/ \
  --data-input-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/passive_1_processed.csv \
  --data-output-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/client1_output.txt \
  --mpc-party-num 3 --sample-num 2115 --class-num 2 \
  --method-id 4 --selected-feature-num 2

./ppdv --party-id 2 --party-num 4 --party-type 1 \
  --network-file ../data/network/Parties-4.txt \
  --log-file ../log/ \
  --data-input-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/passive_2_processed.csv \
  --data-output-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/client2_output.txt \
  --mpc-party-num 3 --sample-num 2115 --class-num 2 \
  --method-id 4 --selected-feature-num 2

./ppdv --party-id 3 --party-num 4 --party-type 1 \
  --network-file ../data/network/Parties-4.txt \
  --log-file ../log/ \
  --data-input-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/passive_3_processed.csv \
  --data-output-file ../data/dataset/rebuttal/bank_scalability/bank_1a3p_1/client3_output.txt \
  --mpc-party-num 3 --sample-num 2115 --class-num 2 \
  --method-id 4 --selected-feature-num 2

party-type 0 is the active party and party-type 1 is a passive party. The C++ clients perform local preprocessing, connect to each other using data/network/Parties-4.txt, and send secret-shared inputs to the MP-SPDZ runtime on port 14000.

Run PrivPSV with one passive party for comparison

Change the MAX_NUM_CLIENTS in Programs/Source/client_input_feature_selection.mpc to 2 (for one active party and one passive party). Compile it in the same way above by using method_id = 2 or method_id = 3.

Run The Single-MI Micro-Benchmark

The micro-benchmark program is:

third_party/MP-SPDZ/Programs/Source/micro_bench_single_mi.mpc

It generates random inputs inside MP-SPDZ and benchmarks one mutual information calculation. It does not require the ppdv C++ clients.

Compile from the MP-SPDZ directory:

cd third_party/MP-SPDZ
./compile.py -R 64 -l Programs/Source/micro_bench_single_mi.mpc -- \
  <sample_num> <nbits> <threads> <method_id>

Arguments:

Argument	Meaning
sample_num	Number of generated samples.
nbits	Bit length of each discretized value. The bucket count is 2^nbits.
threads	Number of MP-SPDZ threads.
method_id	Benchmark method.

Micro-benchmark methods:

Method	Description
1	Sorting-based MI with a naive sequential scan.
2	Sorting-based MI with Blelloch scan parallelization.
3	Indicator-vector-based MI.

Example compile commands:

# Method 1: sorting-based, naive scan
./compile.py -R 64 -l Programs/Source/micro_bench_single_mi.mpc -- 64 3 4 1

# Method 2: sorting-based, Blelloch scan
./compile.py -R 64 -l Programs/Source/micro_bench_single_mi.mpc -- 64 3 4 2

# Method 3: indicator-vector-based
./compile.py -R 64 -l Programs/Source/micro_bench_single_mi.mpc -- 64 3 4 3

Run the benchmark with three MP-SPDZ parties:

cd third_party/MP-SPDZ

./replicated-ring-party.x -I -p 0 micro_bench_single_mi-64-3-4-2
./replicated-ring-party.x -I -p 1 micro_bench_single_mi-64-3-4-2
./replicated-ring-party.x -I -p 2 micro_bench_single_mi-64-3-4-2

Replace the suffix 64-3-4-2 with the arguments used at compile time. The program prints timing markers and the final MI result in the MP-SPDZ output.

Notes

• Keep SPDZ_PORT_BASE in include/common.h consistent with PORT_NUM in the MP-SPDZ program. The current value is 14000.
• Keep SPDZ_FIXED_POINT_PRECISION in include/common.h consistent with FIXED_F in the MP-SPDZ program. The current value is 16.
• MP-SPDZ -l enables flow optimization. It is useful for the current programs, but large datasets can still take a long time to compile.