A sum-product network (SPN) is a graphical model that allows several types of inference to be drawn efficiently. This implementation, provides a privacy-preserving protocol which tackles structure generation and parameter learning of SPNs. To preserve the privacy of the participants, we derive a protocol based on secret sharing, which guarantees privacy in the honest-but-curios setting even when at most half of the parties cooperate to disclose the data. The protocol makes use of a forest of randomly generated SPNs which is trained and weighted privately and can then be used for private inference on data points. Experiments show good log-likelihood performance and the efficiency of the scheme with a varying number of parties on homogeneous and heterogeneous partitioned data.
The manger and members of the network can be started with docker.
- Ensure to install Docker at your machine.
- For running the network locally on your machine create a docker network overlay via
docker network create --driver=overlay --attachable spn_overlay
- Use the dockerfiles for manager and member to build docker images
docker build -t manager_build -f dockerfile_manager .
OR
docker build -t member_build -f dockerfile_member .
- After building the images they can be run via
docker run --rm --name manager --network spn_overlay --ip 10.0.1.10 -e CONFIG_FILE_LOCATION="./resources/config/config_$dataset_$members_docker.ini" manager_build
OR
docker run --rm --name member$ID --network spn_overlay --ip 10.0.1.1$ID -e CONFIG_FILE_LOCATION="./resources/config/config_$dataset_$members_docker.ini" -e ID_OF_MEMBER="$ID" member_build
where $ID is the id of the member you want to run and $dataset the dataset which should be run and $members the number of members in the final network.
- You can close a running docker image with
docker stop manager
OR
docker stop member$ID
Additionally, there are some basic docker commands provided in usefull docker command.txt.
The datasets used can be found at https://github.com/arranger1044/DEBD. For bnetflix, baudio, plants, tretail and nltcs the used config files can be found at resources/config. For those datasets the distributed and differently partitioned data is already included in resources/input for 3, 5 and 10 members, as well as for 3-13 members for the nltcs dataset.
For reproducing the experimental results docker is recommended.
For running the protocol on different dataset partitions and with a varying number of members you have to install the following python packages:
pip install galois==0.0.20 lark matplotlib nest-asyncio networkx scipy sympy torch tqdm pandas==1.3.1 tensorflow==1.15.0 spflow
Python should be at version 3.7.x as otherwise the tensorflow version does not work.
In the src folder of the cloned repository you can start the network with main_$members_$dataset.py where $members is the number of members in the final network and $dataset the dataset on which the protocol should be performed.
Also, run.py or run_5.py can be used to start the protocol for all datasets for 3 or 5 members simultaneously.
For the network to run correct it is important that the IP address of each member is correctly set in all config files.
In this section are settings, used by all members and the manager.
| name | type | meaning |
|---|---|---|
| prim_number | Integer | The prime number used to define the ring over intergers for en- and decryption |
| joint_random_zero_minimum | Integer | The minimal value a share of zero can have as additive share |
| d_multiplyer | Integer | Used to map real values into integers |
| sharing_over_integer_max_secret | Integer | The greatest allowed value to be encrypted as shares over integers |
| sharing_over_integer_allowed_bit_length_of_secret | Integer | The allowed bit-length of a value to be encrypted as shares over integers |
| load_spn_weights | Boolean | If set to true, the network loads the parameters of all SPNs in the forest from the file specified by spn_weights_file_path (should be combined with generate_spn_structures = False) |
| save_spn_weights | Boolean | If set to true, the network saves the parameters of all SPNs in the forest to the file specified by spn_weights_file_path |
| num_dims | Integer | Number of dimensions of the dataset |
| num_local_iterations | Integer | Number of local iterations or stopping criteria for training (20, 30, 40, 100, 300 as number of iterations and 0.001, 0.0005, 0.0001 as stopping criteria are supported) |
| ratspn_dictionary | String | Path to the RAT-SPN dictionary |
| generate_spn_structures | Boolean | If set to true, a new forest is generated, otherwise the forest from the last run is loaded |
| truncate_n | Integer | Number of Newton iterations |
| truncate_t | Integer | Precision of the Newton method |
| private | Boolean | If false, the protocol is performed in a distributed but non-private way |
| output_file_path | String | File were the log-likelihood performance of the protocol is saved, files need to be added when used. |
| sharing_over_integer_security_parameter | Integer | Currently not in use, but left for further improvements |
Each ID section is of pattern ID_x, where x should be identical to the ID in the section itself.
| name | type | meaning |
|---|---|---|
| id | Integer | ID of the member/manager this section is about, used for en- and decryption |
| ip4 | x.x.x.x | The ip4 adress of this member/manager |
| port | Integer | Port used for the communication with the network |
| name | String | The name for this member/manager, not used for calculations and comunications |
| latency | Integer | The delay this member/manager has at sending/receiving a message, with unit milliseconds (ms) |
| private_evaluation | Boolean | If set true, the first x lines of the file specified by private_data_for_evaluation_file_path are evaluated, x is specified by private_evaluation_amount_lines_to_evaluate |
| private_evaluation_amount_lines_to_evaluate | Integer | The number of lines which should be evaluated of the file specified by private_data_for_evaluation_file_path, only invoked when private_evaluation is set to True |
| private_data_file_path | String | The filepath of the private data used for training the parameters of the RAT-SPN forest |
| spn_file_path | String | The filepath of the saved forest, where all SPNs are pickled objects |
| spn_weights_file_path | String | The filepath of the shares this member has on all parameters |
| private_data_for_evaluation_file_path | String | The filepath of the private input used for evaluation, which only happens if private_evaluation is set to true |
| saving_checkpoints_file_path | String | Filepath were checkpoints for the best SPNs are saved |
| global_data_for_evaluation_file_path | String | Filepath were the global data for performance measurement is stored |