Fixing and documenting the CrowdGuard workflow tutorial (securefedera…

…tedai#1289) * Fixing CrowdGuard demo script start-up Signed-off-by: Teodor Parvanov <[email protected]> * Updating the readme with instructions for running cifar10_crowdguard.py Signed-off-by: Teodor Parvanov <[email protected]> * Fixing the PoisoningAttackDemo.ipynb notebook Signed-off-by: Teodor Parvanov <[email protected]> * Additional fixes for readme.md Signed-off-by: Teodor Parvanov <[email protected]> * Installing Workflow API requirements from the root workflow folder Signed-off-by: Teodor Parvanov <[email protected]> --------- Signed-off-by: Teodor Parvanov <[email protected]>
ishant162 · Jan 22, 2025 · cafc558 · cafc558
1 parent 4a1a135
commit cafc558
Show file tree

Hide file tree

Showing 4 changed files with 35 additions and 14 deletions.
diff --git a/openfl-tutorials/experimental/workflow/CrowdGuard/CrowdGuardClientValidation.py b/openfl-tutorials/experimental/workflow/CrowdGuard/CrowdGuardClientValidation.py
@@ -377,7 +377,7 @@ def __prune_poisoned_models(num_layers, total_number_of_clients, own_client_inde
 
                 ac_e = AgglomerativeClustering(n_clusters=2, distance_threshold=None,
                                                compute_full_tree=True,
-                                               affinity="euclidean", memory=None,
+                                               metric="euclidean", memory=None,
                                                connectivity=None,
                                                linkage='single',
                                                compute_distances=True).fit(cluster_input)

diff --git a/openfl-tutorials/experimental/workflow/CrowdGuard/PoisoningAttackDemo.ipynb b/openfl-tutorials/experimental/workflow/CrowdGuard/PoisoningAttackDemo.ipynb
@@ -430,9 +430,8 @@
     "        state_dicts = [model.state_dict() for model in models]\n",
     "        state_dict = new_model.state_dict()\n",
     "        for key in models[1].state_dict():\n",
-    "            state_dict[key] = np.sum(\n",
-    "                [state[key] for state in state_dicts], axis=0\n",
-    "            ) / len(models)\n",
+    "            state_dict[key] = torch.from_numpy(\n",
+    "                np.average([state[key].numpy() for state in state_dicts], axis=0))\n",
     "        new_model.load_state_dict(state_dict)\n",
     "    return new_model\n",
     "\n",
@@ -558,8 +557,7 @@
     "            exclude=[\"private\"],\n",
     "            )\n",
     "\n",
-    "    # @collaborator  # Uncomment if you want ro run on CPU\n",
-    "    @collaborator(num_gpus=1)  # Assuming GPU(s) is available on the machine\n",
+    "    @collaborator\n",
     "    def train(self):\n",
     "        self.collaborator_name = self.input\n",
     "        print(20 * \"#\")\n",
@@ -669,7 +667,7 @@
     "\n",
     "        ac_e = AgglomerativeClustering(n_clusters=2, distance_threshold=None,\n",
     "                                       compute_full_tree=True,\n",
-    "                                       affinity=\"euclidean\", memory=None, connectivity=None,\n",
+    "                                       metric=\"euclidean\", memory=None, connectivity=None,\n",
     "                                       linkage='single',\n",
     "                                       compute_distances=True).fit(binary_votes)\n",
     "        ac_e_labels: list = ac_e.labels_.tolist()\n",

diff --git a/openfl-tutorials/experimental/workflow/CrowdGuard/cifar10_crowdguard.py b/openfl-tutorials/experimental/workflow/CrowdGuard/cifar10_crowdguard.py
@@ -220,9 +220,8 @@ def FedAvg(models):  # NOQA: N802
         state_dicts = [model.state_dict() for model in models]
         state_dict = new_model.state_dict()
         for key in models[1].state_dict():
-            state_dict[key] = np.sum(
-                [state[key] for state in state_dicts], axis=0
-            ) / len(models)
+            state_dict[key] = torch.from_numpy(
+                np.average([state[key].numpy() for state in state_dicts], axis=0))
         new_model.load_state_dict(state_dict)
     return new_model
 
@@ -316,8 +315,7 @@ def start(self):
             exclude=["private"],
         )
 
-    # @collaborator  # Uncomment if you want ro run on CPU
-    @collaborator(num_gpus=1)  # Assuming GPU(s) is available on the machine
+    @collaborator
     def train(self):
         self.collaborator_name = self.input
         print(20 * "#")
@@ -428,7 +426,7 @@ def defend(self, inputs):
 
         ac_e = AgglomerativeClustering(n_clusters=2, distance_threshold=None,
                                        compute_full_tree=True,
-                                       affinity="euclidean", memory=None, connectivity=None,
+                                       metric="euclidean", memory=None, connectivity=None,
                                        linkage='single',
                                        compute_distances=True).fit(binary_votes)
         ac_e_labels: list = ac_e.labels_.tolist()

diff --git a/openfl-tutorials/experimental/workflow/CrowdGuard/readme.md b/openfl-tutorials/experimental/workflow/CrowdGuard/readme.md
@@ -20,4 +20,29 @@ We implemented a simple scaling-based poisoning attack to demonstrate the effect
 
 For the local validation in CrowdGuard, each client uses its local dataset to obtain the hidden layer outputs for each local model. Then it calculates the Euclidean and Cosine Distance, before applying a PCA. Based on the first principal component, CrowdGuard employs several statistical tests to determine whether poisoned models remain and removes the poisoned models using clustering. This process is repeated until no more poisoned models are detected before sending the detected poisoned models to the server. On the server side, the votes of the individual clients are aggregated using a stacked-clustering scheme to prevent malicious clients from manipulating the aggregation process through manipulated votes. The client-side validation as well as the server-side operations, are executed with SGX to prevent privacy attacks.
 
-[1] Rieger, P., Krauß, T., Miettinen, M., Dmitrienko, A., & Sadeghi, A. R. CrowdGuard: Federated Backdoor Detection in Federated Learning. NDSS 2024.
+[1] Rieger, P., Krauß, T., Miettinen, M., Dmitrienko, A., & Sadeghi, A. R. CrowdGuard: Federated Backdoor Detection in Federated Learning. NDSS 2024.
+
+## Running the CIFAR-10 demo script
+The demo script requires a dedicated allocation of at least 18GB of RAM to run without issues.
+
+1) Create a Python virtual environment for better isolation
+```shell
+python -m venv venv
+source venv/bin/activate
+```
+2) Install OpenFL from the latest sources
+```shell
+git clone https://github.com/securefederatedai/openfl.git && cd openfl
+pip install -e .
+```
+3) Install the requirements for Workflow API
+```shell
+cd openfl-tutorials/experimental/workflow
+pip install -r workflow_interface_requirements.txt
+```
+4) Start the training script<br/>
+Note that the number of training rounds can be adjusted via the `--comm_round` parameter:
+```shell
+cd CrowdGuard
+python cifar10_crowdguard.py --comm_round 5
+```