final version

examples/mnist-pytorch-DPSGD/README.rst

@@ -12,6 +12,12 @@ Prerequisites
 -  `Python >=3.8, <=3.12 <https://www.python.org/downloads>`__
 -  `A project in FEDn Studio  <https://fedn.scaleoutsystems.com/signup>`__   
 
+Edit Differential Privacy budget
+--------------------------
+- The **Differential Privacy budget** (`FINAL_EPSILON`, `DELTA`) is configured in the `compute` package at `client/train.py` (lines 35 and 39).
+- If `HARDLIMIT` (line 40) is set to `True`, the `FINAL_EPSILON` will not exceed its specified limit.
+- If `HARDLIMIT` is set to `False`, the expected `FINAL_EPSILON` will be around its specified value given the server runs `ROUNDS` variable (line 36).
+
 Creating the compute package and seed model
 -------------------------------------------
 
@@ -26,7 +32,7 @@ Clone this repository, then locate into this directory:
 .. code-block::
 
    git clone https://github.com/scaleoutsystems/fedn.git
-   cd fedn/examples/mnist-pytorch
+   cd fedn/examples/mnist-pytorch-DPSGD
 
 Create the compute package:
 

examples/mnist-pytorch-DPSGD/client/train.py

@@ -37,6 +37,7 @@ def __getitem__(self, idx):
 EPOCHS = 5
 EPSILON = FINAL_EPSILON/ROUNDS
 DELTA = 1e-5
+HARDLIMIT = False
 
 MAX_PHYSICAL_BATCH_SIZE = 32
 
@@ -111,24 +112,28 @@ def train(in_model_path, out_model_path, data_path=None, batch_size=32, epochs=1
 
     d_epsilon = privacy_engine.get_epsilon(DELTA)
     print("epsilon spent: ", d_epsilon)
-    tot_epsilon += d_epsilon
+    tot_epsilon = np.sqrt(tot_epsilon**2 d_epsilon**2)
     print("saving tot_epsilon: ", tot_epsilon)
     np.save('epsilon.npy', tot_epsilon)
 
-    # Metadata needed for aggregation server side
-    metadata = {
-        # num_examples are mandatory
-        "num_examples": len(x_train),
-        "batch_size": batch_size,
-        "epochs": epochs,
-        "lr": lr,
-    }
+    if HARDLIMIT and tot_epsilon >= FINAL_EPSILON:
+        print("DP Budget Exceeded: The differential privacy budget has been exhausted, no model updates will be applied to preserve privacy guarantees.")
 
-    # Save JSON metadata file (mandatory)
-    save_metadata(metadata, out_model_path)
-
-    # Save model update (mandatory)
-    save_parameters(model, out_model_path)
+    else:
+        # Metadata needed for aggregation server side
+        metadata = {
+            # num_examples are mandatory
+            "num_examples": len(x_train),
+            "batch_size": batch_size,
+            "epochs": epochs,
+            "lr": lr,
+        }
+
+        # Save JSON metadata file (mandatory)
+        save_metadata(metadata, out_model_path)
+
+        # Save model update (mandatory)
+        save_parameters(model, out_model_path)
 
 def accuracy(preds, labels):
     return (preds == labels).mean()