TraceFL: Interpretability-Driven Debugging in Federated Learning via Neuron Provenance

Accepted at 2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE) [Arxiv Paper Link]

For questions or feedback, please contact at waris@vt.edu. The code is written in Flower FL Framework, the most widely used FL framework.

1. TraceFL

TraceFL is a tool designed to provide interpretability in Federated Learning (FL) by identifying clients responsible for specific predictions made by a global model.

1.1 Overview

Federated Learning (FL) enables multiple clients (e.g., hospitals ) to collaboratively train a global model without sharing their raw data. However, this distributed and privacy-preserving setup makes it challenging to attribute a model's predictions to specific clients. Understanding which clients are most responsible for a model's output is crucial for debugging, accountability, and incentivizing high-quality contributions.

TraceFL addresses this challenge by dynamically tracking the significance of neurons in a global model's prediction and mapping them back to the corresponding neurons in each participating client's model. This process allows FL developers to localize the clients most responsible for a prediction without accessing their raw training data.

1.2 Key Features

• Neuron Provenance: A novel technique that tracks the flow of information from individual clients to the global model, identifying the most influential clients for each prediction.
• High Accuracy: TraceFL achieves 99% accuracy in localizing responsible clients in both image and text classification tasks.
• Wide Applicability: Supports multiple neural network architectures, including CNNs (e.g., ResNet, DenseNet) and any transformers model from HuggingFace library (e.g., BERT, GPT).
• Scalability and Robustness: Efficiently scales to thousands of clients and maintains high accuracy under varying data distributions and differential privacy settings.
• No Client-Side Instrumentation Required: Runs entirely on the central server, without needing access to clients' training data or modifications to the underlying fusion algorithm.

2. Running TraceFL

The .sh (e.g., job_training_all_exps.sh) scripts and TraceFL/tracefl/conf/base.yaml provided in this artifact can be used to regenerate any experiment results presented in the paper. `

The experiments cover various aspects of federated learning, including:

1. Image and Text Classification: Evaluating the performance of different models and datasets in federated settings.
2. Differential Privacy: Analyzing the impact of differential privacy on model training and TraceFL's localizability.
3. Scalability: Testing the scalability of TraceFL with varying numbers of clients and rounds.
4. Dirichlet Alpha Tuning: Exploring the effects of different Dirichlet alpha values on data distribution, TraceFL's localizability, and model performance.

2.1 Experiments Configuration Overview

• Image Classification:
  • Models: ResNet18, DenseNet121
  • Datasets: MNIST, CIFAR-10, PathMNIST, OrganAMNIST
  • Number of Rounds: 25-50
• Text Classification:
  • Models: OpenAI GPT, Google BERT
  • Datasets: DBPedia, Yahoo Answers
  • Number of Rounds: 25

2.2 Differential Privacy Analysis

These experiments evaluate the impact of differential privacy on TraceFL by applying different noise levels and clipping norms.

• Models: DenseNet121, OpenAI GPT
• Datasets: MNIST, PathMNIST, DBPedia
• Noise Levels: 0.0001, 0.0003, 0.0007, 0.0009, 0.001, 0.003
• Clipping Norms: 15, 50
• Number of Rounds: 15

2.3 Scalability Experiments

Scalability tests involve running experiments with varying numbers of clients and rounds to assess how well TraceFL scales.

• Models: OpenAI GPT
• Dataset: DBPedia
• Number of Clients: 200, 400, 600, 800, 1000
• Clients per Round: 10, 20, 30, 40, 50
• Number of Rounds: 15, 100

2.4 Dirichlet Alpha Experiments

These experiments explore the effect of different Dirichlet alpha values on data partitioning, model training, and TraceFL's localizability.

• Models: OpenAI GPT, DenseNet121
• Datasets: Yahoo Answers, DBPedia, PathMNIST, OrganAMNIST, MNIST, CIFAR-10
• Dirichlet Alpha Values: 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1
• Number of Clients: 100
• Clients per Round: 10
• Number of Rounds: 15

2.5 Results and Log Files

Each experiment's output will be logged in the logs directory, providing detailed information about the training process and results.

3. Potential Use Cases of TraceFL

• Debugging and Fault Localization: Identify and isolate faulty or malicious clients responsible for incorrect or suspicious predictions in federated learning models.
• Enhancing Model Quality, Fairness, and Incentivization: Improve model performance by rewarding high-quality clients, ensuring fair client contributions, and incentivizing continued participation from beneficial clients.
• Client Accountability and Security: Increase accountability by tracing model decisions back to specific clients, deterring malicious behavior, and ensuring secure contributions.
• Optimized Client Selection and Efficiency: Dynamically select the most beneficial clients for training to enhance model performance and reduce communication overhead.
• Interpretable Federated Learning in Sensitive Domains: Provide transparency and interpretability in federated learning models, crucial for compliance, trust, and ethical considerations in domains like healthcare and finance.

4. Citation

Latex

@inproceedings{gill2025tracefl,
  title = {{TraceFL: Interpretability-Driven Debugging in Federated Learning via Neuron Provenance}},
  author = {Gill, Waris and Anwar, Ali and Gulzar, Muhammad Ali},
  booktitle = {2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE)},
  year = {2025},
  organization = {IEEE},
}