use roc vertical averaging in HFL to save communication overhead (#518)

python/primihub/FL/logistic_regression/hfl_client.py

@@ -279,7 +279,7 @@ class Plaintext_Client:
                                                      drop_intermediate=False)
             self.server_channel.send("fpr", fpr)
             self.server_channel.send("tpr", tpr)
-            self.server_channel.send("thresholds", thresholds)
+            # self.server_channel.send("thresholds", thresholds)
 
             # ks
             ks = ks_from_fpr_tpr(fpr, tpr)

python/primihub/FL/logistic_regression/hfl_server.py

@@ -6,7 +6,7 @@ from primihub.utils.logger_util import logger
 import json
 import numpy as np
 from phe import paillier
-from primihub.FL.metrics.hfl_metrics import fpr_tpr_merge2,\
+from primihub.FL.metrics.hfl_metrics import roc_vertical_avg,\
                                             ks_from_fpr_tpr,\
                                             auc_from_fpr_tpr
 from .base import PaillierFunc
@@ -157,21 +157,15 @@ class Plaintext_DPSGD_Server:
     def get_fpr_tpr(self):
         client_fpr = self.client_channel.recv_all('fpr')
         client_tpr = self.client_channel.recv_all('tpr')
-        client_thresholds = self.client_channel.recv_all('thresholds')
+        # client_thresholds = self.client_channel.recv_all('thresholds')
 
         # fpr & tpr
-        # Note: fpr_tpr_merge2 only support two clients
-        #       use ROC averaging when for multiple clients
+        # roc_vertical_avg: sample = 0.1 * n
+        samples = int(0.1 * sum(self.num_examples_weights))
         fpr,\
-        tpr,\
-        thresholds = fpr_tpr_merge2(client_fpr[0],
-                                    client_tpr[0],
-                                    client_thresholds[0],
-                                    client_fpr[1],
-                                    client_tpr[1],
-                                    client_thresholds[1],
-                                    self.num_positive_examples_weights,
-                                    self.num_negtive_examples_weights)
+        tpr = roc_vertical_avg(samples,
+                               client_fpr,
+                               client_tpr)
         return fpr, tpr
 
     def get_metrics(self):

python/primihub/FL/neural_network/hfl_client.py

@@ -317,7 +317,7 @@ class Plaintext_Client:
                                                      drop_intermediate=False)
             self.server_channel.send("fpr", fpr)
             self.server_channel.send("tpr", tpr)
-            self.server_channel.send("thresholds", thresholds)
+            # self.server_channel.send("thresholds", thresholds)
 
             client_metrics['train_fpr'] = fpr
             client_metrics['train_tpr'] = tpr

python/primihub/FL/neural_network/hfl_server.py

@@ -6,9 +6,9 @@ from primihub.utils.logger_util import logger
 import json
 import numpy as np
 import torch
-from primihub.FL.metrics.hfl_metrics import fpr_tpr_merge2,\
-                                                      ks_from_fpr_tpr,\
-                                                      auc_from_fpr_tpr
+from primihub.FL.metrics.hfl_metrics import roc_vertical_avg,\
+                                            ks_from_fpr_tpr,\
+                                            auc_from_fpr_tpr
 from .base import create_model
 
 
@@ -203,21 +203,15 @@ class Plaintext_Server:
     def get_fpr_tpr(self):
         client_fpr = self.client_channel.recv_all('fpr')
         client_tpr = self.client_channel.recv_all('tpr')
-        client_thresholds = self.client_channel.recv_all('thresholds')
+        # client_thresholds = self.client_channel.recv_all('thresholds')
 
         # fpr & tpr
-        # Note: fpr_tpr_merge2 only support two clients
-        #       use ROC averaging when for multiple clients
+        # roc_vertical_avg: sample = 0.1 * n
+        samples = int(0.1 * sum(self.num_examples_weights))
         fpr,\
-        tpr,\
-        thresholds = fpr_tpr_merge2(client_fpr[0],
-                                    client_tpr[0],
-                                    client_thresholds[0],
-                                    client_fpr[1],
-                                    client_tpr[1],
-                                    client_thresholds[1],
-                                    self.num_positive_examples_weights,
-                                    self.num_negtive_examples_weights)
+        tpr = roc_vertical_avg(samples,
+                               client_fpr,
+                               client_tpr)
         return fpr, tpr
 
     def get_metrics(self):