💻 computer science

Auditable Federated Learning for Multimodal Brain Tumor Classification under Label Skew and Client Dropout

This study proposes an auditable federated learning framework that maintains high accuracy and F1-scores for multimodal brain tumor classification under realistic clinical stressors like label skew and client dropout, while introducing round-level weight auditing to provide forensic visibility into model evolution that conventional metrics miss.

Original authors: Thiruvenkadam T, Sivakumar N, Emerson Raja

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thiruvenkadam T, Sivakumar N, Emerson Raja

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Hospitals sit on vast reservoirs of medical images, from the quick snapshots of computed tomography scans to the detailed, soft-tissue views of magnetic resonance imaging. These pictures hold the key to diagnosing brain tumors early, a critical step that can save lives. Yet, for decades, these images have remained locked inside individual hospitals. Strict privacy laws and ethical concerns prevent institutions from sharing raw patient data, even when pooling that data together could train smarter, more accurate artificial intelligence. To solve this, researchers developed a method called federated learning. Instead of moving the sensitive images to a central computer, this approach sends the learning algorithm to the hospital. The local computer trains on the data it holds, learns the patterns, and sends only the mathematical lessons back to a central server. The server combines these lessons to build a smarter global model, which is then sent back out for the next round of learning. This allows many hospitals to collaborate on a single, powerful diagnostic tool without ever exchanging a single patient photo.

However, the real world is rarely as tidy as the theory. In a perfect scenario, every hospital would have a similar mix of healthy and sick patients, and every computer would be online and ready to learn at the exact same time. In reality, one hospital might see mostly healthy brains while another sees mostly tumors, and a hospital's computer might go offline due to maintenance or network issues. These mismatches and interruptions can confuse the learning process, causing the final model to become less accurate or biased. Furthermore, while the system is designed to be private, it has historically been a "black box" to regulators and doctors. They could see the final accuracy score, but they had no way to inspect the invisible steps the model took to get there, leaving them unable to verify if the model was behaving correctly under stress.

A team of researchers has now built a new system that brings transparency to this process while testing its limits under realistic pressure. They focused on the difficult task of distinguishing between healthy brains and those with tumors, using a combined dataset of nearly ten thousand images from both CT and MRI scans. To simulate a real network of five different hospitals, they created two distinct scenarios. In the first, they distributed the images evenly so every "hospital" saw a balanced mix of healthy and tumor cases. In the second, they created a much harsher environment where each hospital was heavily skewed, seeing mostly one type of case and very few of the other. They also introduced a layer of chaos by randomly disconnecting one hospital about thirty percent of the time, mimicking the sporadic availability found in actual clinical networks.

The researchers trained a compact, efficient neural network—a type of computer program designed to recognize patterns in images—using a specific technique called FedProx. This method helps the local computers stay aligned with the global goal even when their local data is unbalanced. As the training progressed over ten rounds of communication, the system did something new: it kept a detailed, round-by-round log of the model's internal weights. These weights are the adjustable numbers inside the program that determine how it interprets the images. The researchers recorded the state of these numbers before training, after the local updates, and after the global combination. This created a complete audit trail, allowing them to watch the model evolve in real-time.

The results showed that the system was remarkably resilient. Even when the data was heavily skewed and hospitals dropped out frequently, the model maintained high performance. In the balanced scenario, the final model correctly identified tumors and healthy brains about ninety-six percent of the time. In the difficult, skewed scenario, it still achieved nearly ninety-five percent accuracy. The drop in performance was minimal, suggesting that the method could handle the messy reality of hospital data without breaking down. Perhaps more importantly, the audit trail revealed a hidden story that the standard accuracy scores missed. While the final accuracy looked similar in both scenarios, the internal numbers of the model told a different tale. In the skewed scenario, the average value of the final layer's weights shifted consistently in one direction, a subtle signature of the data imbalance that remained even after the model learned to classify correctly.

This discovery suggests that looking at the internal numbers of a federated model can act as a forensic tool. Just as a mechanic might listen to the sound of an engine to detect a problem that a speedometer doesn't show, doctors and regulators can inspect these weight summaries to detect if the training data was unbalanced or if the model is drifting. The researchers found that this shift was invisible in the standard performance charts but clearly visible in the weight logs. This means that hospitals can now verify the integrity of their collaborative AI without needing to see the private patient data. The study demonstrates that it is possible to build a trustworthy, transparent, and robust system for brain tumor detection that works even when data is uneven and participants are sporadically available, offering a practical path forward for deploying medical AI in the real world.

Technical Summary: Auditable Federated Learning for Multimodal Brain Tumor Classification

Problem Statement
Federated Learning (FL) offers a privacy-preserving mechanism for hospitals to collaboratively train deep learning models without exchanging raw patient data. However, existing FL studies in medical imaging often rely on unrealistic assumptions: that client data is Independent and Identically Distributed (IID) and that all clients participate in every communication round. In real-world clinical networks, data heterogeneity (non-IID) due to varying patient demographics and imaging protocols is common, and client dropout due to network issues or maintenance is frequent. Furthermore, while FL relies on the iterative averaging of client weight updates, the optimization dynamics of these weights are rarely audited at the round level. This lack of transparency limits the ability of clinicians and regulators to inspect how global models evolve under stress or to detect distributional drift.

Methodology
The study proposes an auditable federated framework for binary brain tumor classification using a combined multimodal dataset of 4,618 CT slices and 5,000 MRI slices (9,618 total images). The methodology involves four key components:

Model Architecture: A lightweight, three-block Convolutional Neural Network (CNN) is employed to ensure feasibility in resource-constrained hospital settings. The architecture includes three convolutional blocks (32, 64, and 128 filters) followed by dense layers, totaling a compact parameter count suitable for edge deployment.
Federated Optimization: The framework utilizes FedProx with a proximal coefficient ( $\mu$ ) of 0.01 to mitigate client drift caused by non-IID data, combined with FedAvg for global aggregation.
Experimental Stress Testing: The system is evaluated across five simulated clients under two distinct partitioning scenarios:
- IID: Balanced distribution of healthy and tumor cases per client.
- Non-IID: An aggressive label-skew scenario where specific clients are assigned a majority class. Clients 1 and 3 receive a healthy majority (923 healthy vs. 361 tumor samples, ~72% healthy), while clients 2 and 4 receive a tumor majority (230 healthy vs. 1054 tumor samples, ~82% tumor). Client 5 receives the remainder.
- Client Dropout: Approximately 30% of communication rounds involve the random exclusion of one client to simulate intermittent availability.
Weight Auditing Protocol: A core innovation of the framework is the round-level logging of weight statistics. For every active client in every round, the system records three weight summaries:
- Initial: Global model weights received from the server.
- Updated: Local model weights after training.
- Aggregated: New global model weights after aggregation.
  These summaries include per-layer mean, standard deviation, minimum, and maximum values, alongside standard performance metrics (accuracy, loss, F1-score) and deployment metrics (round time, communication payload, memory).

Key Contributions
The paper outlines four primary contributions:

Auditable FL Protocol: A lightweight protocol that tracks initial, updated, and aggregated weight statistics (mean, std, min, max) for each active client at every round. This trace is compatible with privacy-preserving aggregation as it does not require sharing raw gradients or data.
Realistic Stress Testing: Evaluation on a large multimodal dataset (9,618 images) under severe non-IID conditions (aggressive label skew) and intermittent client dropout (~30% of rounds), moving beyond idealized IID/full-participation benchmarks.
Deployment-Aware Metrics: Reporting of practical indicators including wall-clock round time, communication payload, and process memory, enabling assessment of both predictive quality and operational cost.
Weight-Audit as a Diagnostic Signal: Empirical demonstration that weight-audit traces reveal a measurable bias signature in the output-layer mean during non-IID training. This shift is invisible in conventional aggregate accuracy curves but serves as a forensic signal of distributional drift.

Results

Performance: The global model achieved 96.05% accuracy and an F1-score of 0.9641 under the IID partition. Under the severe non-IID partition with 30% client dropout, it achieved 94.96% accuracy and an F1-score of 0.9547. This represents a degradation of only 1.09 percentage points, demonstrating robustness against class skew and client dropout.
Convergence: The accuracy gap between IID and non-IID scenarios narrowed from 3.43 percentage points in round 1 to 1.09 percentage points by round 10, indicating that FedProx and uniform aggregation effectively absorbed per-client skew over time.
Weight Audit Findings: Analysis of the output-layer weights revealed a persistent, measurable shift in the mean under non-IID training. By round 10, the output-layer mean for the non-IID scenario was approximately 2.04 times higher than the IID scenario (0.0418 vs. 0.0205). This bias signature was detectable despite the models converging to similar accuracy levels.
Deployment Metrics: The framework maintained consistent performance costs, with average round times of ~45 seconds and communication payloads of ~32 MB per round, regardless of the partitioning scenario.

Significance and Claims
The paper argues that explicit per-round weight reporting should become a standard practice in federated medical imaging research to enhance scientific transparency and clinical trust. The primary significance lies in the demonstration that round-level weight inspection serves as a lightweight forensic signal for federated medical AI. Specifically, the study claims that:

Weight-audit traces can expose distributional bias and drift that are concealed by standard accuracy and loss curves.
This auditing mechanism requires no additional model parameters, adds negligible computational cost, and preserves data privacy.
The framework provides a practical, low-cost integrity check for regulators and clinicians to verify that model updates remain within expected statistical bounds, even under the stress of non-IID data and client dropout.

The authors maintain a modest stance, acknowledging limitations such as the use of a single dataset, the lack of formal statistical hypothesis testing on the bias signature, and the absence of explainable AI (XAI) techniques like Grad-CAM in the current iteration. They position the work as a step toward more transparent and trustworthy federated systems rather than a final solution for all medical AI challenges.

Technical Summary: Auditable Federated Learning for Multimodal Brain Tumor Classification

More like this