app(flowerhub): add FedXGBoost for financial fraud detection implementation by eo4929 · Pull Request #6807 · flwrlabs/flower

eo4929 · 2026-03-20T16:12:55Z

Summary

Implement FedXGBoost algorithm
Add training and evaluation pipeline
Ensure compatibility with Flower framework

Validation

./framework/dev/format.sh
./framework/dev/test.sh

I follow some instructions written in https://www.notion.so/flowerlabs/Guide-How-to-Publish-Apps-on-Flower-Hub-1d1d8ccd59cf8073a742da2c83ceb89b.

…dating .toml)

Copilot

Pull request overview

Adds a new Flower Hub app implementing a federated XGBoost-based workflow for financial fraud detection, including client/server apps, data utilities, and ensemble aggregation.

Changes:

Introduces a Flower ServerApp/ClientApp training + evaluation loop for “FedXGBBagging”.
Adds dataset preprocessing, partitioning utilities, and XGBoost model (de)serialization helpers.
Adds a large fed_xgb_bagging.py module implementing bagging and similarity-based aggregation utilities.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
examples/FinancialFraudDetection-app/pyproject.toml	Declares the app package metadata, dependencies, and Flower app entrypoints/config.
examples/FinancialFraudDetection-app/frauddetection/task.py	Implements preprocessing, data loading/partitioning, training, evaluation, and model serialization helpers.
examples/FinancialFraudDetection-app/frauddetection/server_app.py	Implements the federated orchestration loop, collects per-round client models, builds an ensemble, and runs evaluation.
examples/FinancialFraudDetection-app/frauddetection/client_app.py	Implements per-client local training and evaluation handlers and transmits serialized boosters.
examples/FinancialFraudDetection-app/frauddetection/fed_xgb_bagging.py	Adds ensemble/bagging and similarity-based utilities used server-side for prediction/evaluation.
examples/FinancialFraudDetection-app/frauddetection/init.py	Adds package marker and module docstring.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-03-20T16:16:04Z

examples/FinancialFraudDetection-app/pyproject.toml

+[project]
+name = "federated-fraud-detection"
+version = "1.0.0"
+description = "Federated Financial Fraud Detection with XGBoost and Flower"
+license = "Apache-2.0"
+dependencies = [
+    "flwr[simulation]>=1.26.1",
+    "xgboost>=2.0.0",
+    "scikit-learn>=1.3.0",
+    "pandas>=2.0.0",
+    "numpy>=1.24.0",
+]
+
+[tool.hatch.build.targets.wheel]
+packages = ["."]


packages = [\".\"] is very likely to produce an incorrect wheel (it may include unintended files and/or fail to package frauddetection as an importable package). Define the actual package directory (e.g., frauddetection) via Hatch's package selection, and consider adding requires-python since the code uses modern typing syntax (e.g., dict | None in task.py) which requires Python 3.10+.

Copilot · 2026-03-20T16:16:05Z

examples/FinancialFraudDetection-app/frauddetection/task.py

@@ -0,0 +1,191 @@
+"""frauddetection: XGBoost model training and data utilities."""
+
+import json


json is imported but not used in this module. Removing it avoids dead imports and keeps the module lint-clean.

Suggested change

import json

Copilot · 2026-03-20T16:16:05Z

examples/FinancialFraudDetection-app/frauddetection/task.py

+    df = pd.read_csv(data_csv)
+    df = df.sample(frac=1, random_state=42).reset_index(drop=True)
+
+    n = len(df)
+    size = n // num_partitions
+    start = partition_id * size
+    end = start + size if partition_id < num_partitions - 1 else n
+    partition = df.iloc[start:end].reset_index(drop=True)
+
+    X, y = preprocess_df(partition)
+    X_train, X_test, y_train, y_test = _split(X, y)
+    return X_train, X_test, y_train, y_test


If len(df) < num_partitions, then size becomes 0 and most partitions will be empty (partition has 0 rows). That will cause train_test_split to throw (no samples) or produce invalid behavior. Consider handling size == 0 explicitly (e.g., cap num_partitions to n, or compute split indices using np.array_split, or raise a clear error).

Copilot · 2026-03-20T16:16:05Z

examples/FinancialFraudDetection-app/frauddetection/server_app.py

+        # Send the *first* collected model to ``fraction_evaluate`` of clients
+        # as a representative model so they can report per-partition metrics.
+        n_eval = max(1, int(fraction_evaluate * n_clients))
+        eval_node_ids = node_ids[:n_eval]
+
+        # Load the first model as the "representative" model for evaluation
+        with open(all_model_paths[0], "rb") as f:
+            rep_model_bytes = f.read()
+        rep_model_array = np.frombuffer(rep_model_bytes, dtype=np.uint8).copy()


This will crash with IndexError when all_model_paths is empty (e.g., n_clients == 0, no replies, or if model collection fails earlier). Add an explicit guard before indexing and either skip evaluation or raise a clear runtime error when no models were collected.

Copilot · 2026-03-20T16:16:05Z

examples/FinancialFraudDetection-app/frauddetection/server_app.py

+            print("\n[Server] Central evaluation on held-out server data …")
+            X_test, y_test = _load_central_test(csv_path)
+            ensemble.test_data = X_test
+            y_pred, y_prob = ensemble.predict(X_test, y_true=y_test)


Passing y_true=y_test into ensemble.predict (which auto-selects a best-F1 threshold) tunes the decision threshold on the same held-out set you then report metrics on, inflating results (data leakage). For reporting central test metrics, use a fixed threshold (e.g., 0.5) or select the threshold on a separate validation split.

Suggested change

y_pred, y_prob = ensemble.predict(X_test, y_true=y_test)

y_pred, y_prob = ensemble.predict(X_test)

Copilot · 2026-03-20T16:16:06Z

examples/FinancialFraudDetection-app/frauddetection/fed_xgb_bagging.py

+        self.config = config
+        self.result_path = result_path
+        self.history = []
+        self.test_data = None
+
+        self.models = [self._load_model(path) for path in model_paths]
+
+        logger.info(f"Initialized EachBankModel for bank_name_round_number={self.config['bank_name_round_number']}")


config defaults to None, but the constructor immediately subscripts it (self.config['bank_name_round_number']), which will raise TypeError when config is None and KeyError when the key is missing. Make config always a dict (e.g., self.config = config or {}) and use .get(...) with defaults in logs/filenames to keep the class usable as a general utility.

Suggested change

self.config = config

self.result_path = result_path

self.history = []

self.test_data = None

self.models = [self._load_model(path) for path in model_paths]

logger.info(f"Initialized EachBankModel for bank_name_round_number={self.config['bank_name_round_number']}")

# Ensure config is always a dictionary to avoid TypeError/KeyError when accessing it.

self.config = config or {}

self.result_path = result_path

self.history = []

self.test_data = None

self.models = [self._load_model(path) for path in model_paths]

bank_name_round_number = self.config.get("bank_name_round_number", "unknown")

logger.info(f"Initialized EachBankModel for bank_name_round_number={bank_name_round_number}")

Copilot · 2026-03-20T16:16:06Z

examples/FinancialFraudDetection-app/frauddetection/fed_xgb_bagging.py

+    def save_metrics_history(self, metrics, detection_summary, test_time=None, output_path=None):
+        logger.info("Saving metrics and detection summary to history")
+
+        def convert_to_serializable(obj):
+            if isinstance(obj, (np.generic, np.int64, np.float32)):
+                return obj.item()
+            elif isinstance(obj, np.ndarray):
+                return obj.tolist()
+            elif isinstance(obj, pd.Series):
+                return obj.to_dict()
+            elif isinstance(obj, pd.DataFrame):
+                return obj.to_dict(orient="index")
+            return obj
+
+        def flatten_grouped_stats(grouped_stats_df: pd.DataFrame) -> dict:
+            result = {}
+            for group in grouped_stats_df.index:
+                stats_dict = {}
+                for (feature, stat), value in grouped_stats_df.loc[group].items():
+                    stats_dict[f"{feature}__{stat}"] = value
+                result[group] = stats_dict
+            return result
+
+        metrics_path = output_path / "metrics.json"
+        with open(metrics_path, "w") as f:
+            json.dump(metrics, f, indent=4, default=convert_to_serializable)


output_path defaults to None but is treated as a Path (output_path / \"metrics.json\"), which will throw at runtime if the caller doesn't pass it. Either make output_path a required Path parameter (preferred), or add a guard/auto-create a default directory before using it.

Copilot · 2026-03-20T16:16:06Z

examples/FinancialFraudDetection-app/frauddetection/fed_xgb_bagging.py

+
+def _metadata_similarity(meta_i: Dict, meta_j: Dict) -> float:
+
+    score_parts: List[Tuple[str, float, float]] = []  # (ì´ë¦, ì ì, ê°ì¤ì¹)


There are multiple comments/log lines in this module that appear to be mojibake (garbled encoding) and/or non-English in a way that’s not readable in the repository’s context. Please normalize these to clear, UTF-8, developer-facing English comments/log messages so future maintainers can understand intent (and to avoid tooling/display issues).

Suggested change

score_parts: List[Tuple[str, float, float]] = [] # (ì�´ë¦�, ì �ì��, ê°�ì¤�ì¹�)

score_parts: List[Tuple[str, float, float]] = [] # (name, score, weight)

Copilot · 2026-03-20T16:16:07Z

examples/FinancialFraudDetection-app/frauddetection/client_app.py

+import os
+


os is imported but not used in this file. Remove it to keep the module clean and avoid lint failures.

Suggested change

import os

eo4929 added 2 commits March 21, 2026 00:51

feat: add FedXGBoost for financial fraud detection implementation

38707a3

feat: add FedXGBoost for financial fraud detection implementation (up…

70abf77

…dating .toml)

eo4929 requested review from chongshenng, danieljanes, jafermarq, panh99, tanertopal and yan-gao-GY as code owners March 20, 2026 16:12

Copilot AI review requested due to automatic review settings March 20, 2026 16:12

Copilot started reviewing on behalf of eo4929 March 20, 2026 16:13 View session

Copilot AI reviewed Mar 20, 2026

View reviewed changes

github-actions bot added the Contributor Used to determine what PRs (mainly) come from external contributors. label Mar 20, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

app(flowerhub): add FedXGBoost for financial fraud detection implementation#6807

app(flowerhub): add FedXGBoost for financial fraud detection implementation#6807
eo4929 wants to merge 2 commits intoflwrlabs:mainfrom
eo4929:FedXGBoost

eo4929 commented Mar 20, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Copilot AI Mar 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		@@ -0,0 +1,191 @@
		"""frauddetection: XGBoost model training and data utilities."""

		import json

	y_pred, y_prob = ensemble.predict(X_test, y_true=y_test)
	y_pred, y_prob = ensemble.predict(X_test)


		def _metadata_similarity(meta_i: Dict, meta_j: Dict) -> float:

		score_parts: List[Tuple[str, float, float]] = [] # (ì´ë¦, ì ì, ê°ì¤ì¹)

Conversation

eo4929 commented Mar 20, 2026

Summary

Validation

I follow some instructions written in https://www.notion.so/flowerlabs/Guide-How-to-Publish-Apps-on-Flower-Hub-1d1d8ccd59cf8073a742da2c83ceb89b.

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants