💻 computer science

Communication-Efficient FedProx-Binary FedVote for Dual-Task Wi-Fi Fingerprint Indoor Localization

This paper proposes a communication-efficient FedProx-Binary FedVote framework for dual-task Wi-Fi fingerprint indoor localization that significantly reduces uplink bandwidth by transmitting 1-bit update directions while maintaining high accuracy and robustness under strong Non-IID data conditions through FedProx regularization.

Original authors: Guogen Yuan, Zongchang Yang

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guogen Yuan, Zongchang Yang

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

In the quiet hum of a shopping mall, a hospital corridor, or a university campus, invisible signals from Wi-Fi routers are constantly bouncing off walls, furniture, and people. These signals create a unique, shifting pattern of strength and weakness at every specific spot, much like a fingerprint left behind by the air itself. For decades, engineers have tried to use these patterns to tell a phone exactly where it is inside a building, a task that global positioning systems cannot do because the sky is blocked by the roof. The traditional way to build a system that understands these patterns involves gathering thousands of these signal samples from different people and sending them all to a central computer. However, this approach raises two significant problems: it consumes a lot of internet bandwidth to send the data, and it requires people to share their private movement history, which many are unwilling to do.

To solve this, researchers have turned to a method called federated learning, where the computer model travels to the data instead of the data traveling to the computer. In this setup, each phone learns from its own local signal patterns and sends only the lessons it learned back to a central server, keeping the raw data safe on the device. While this protects privacy, it creates a new bottleneck: the phones still have to send large, detailed summaries of their lessons, which can clog the network. A team of researchers from Hunan University of Science and Technology has now developed a way to make this process much lighter and more robust, allowing phones to send their lessons as simple, one-bit directions rather than complex, heavy data files, even when the data on each phone is very different from the others.

The researchers focused on a specific challenge in indoor navigation: knowing not just the general area, but the exact building and the specific floor. In large complexes like universities or hospitals, a system must correctly identify both the building and the floor to be useful. The team set up an experiment using a real-world dataset of Wi-Fi signals collected from multiple buildings and floors. They simulated a network of ten different users, each holding a phone with a unique set of signal data. In the real world, one user might spend most of their time on the first floor of a library, while another is always on the third floor of a science building. This uneven distribution of data, known as non-IID, makes it difficult for a central computer to agree on a single, accurate map because every user is trying to teach the system a slightly different version of reality.

The team first tested a standard method where phones sent full, detailed updates to the server. This worked well when the data was somewhat balanced, but it required sending about 200.33 megabytes of data over the network. They then tried a new approach inspired by a technique called FedVote, where instead of sending the full lesson, each phone simply sent a single bit of information indicating whether a specific part of the model should be adjusted up or down. This reduced the data sent to the server to just 6.26 megabytes, a massive reduction that kept the network light. Under conditions where the data was moderately uneven, this simple method performed just as well as the heavy, detailed method, correctly identifying the building and floor in nearly 91 percent of cases.

However, the researchers found that when the data was extremely uneven—meaning each user was stuck in a very specific, isolated part of the building—the simple one-bit method began to stumble. The phones were sending conflicting directions, and the central server struggled to find a clear path forward. To fix this, the team added a stabilizing rule to the local training process. This rule gently encouraged each phone to stay close to the current global model, preventing it from drifting too far into its own narrow perspective. By combining this stabilizing rule with the one-bit voting system, the researchers created a method that could handle even the most difficult, uneven data scenarios.

The results showed that this combined approach was highly effective. When the data was heavily skewed, the simple voting method alone managed to get the building and floor right about 69.76 percent of the time. But with the added stabilizing rule, the accuracy jumped to 86.91 percent, all while keeping the data transmission at the tiny 6.26 megabyte level. The researchers demonstrated that it is possible to achieve high-precision indoor localization without flooding the network with data or compromising user privacy. Their work suggests that future indoor navigation systems can be both efficient and accurate, relying on simple, lightweight signals from many devices to build a clear picture of a complex indoor world, without ever needing to see the raw data itself.

Technical Summary: Communication-Efficient FedProx-Binary FedVote for Dual-Task Wi-Fi Fingerprint Indoor Localization

1. Problem Statement

Wi-Fi fingerprinting is a widely used solution for indoor localization as it leverages existing access points without requiring additional infrastructure. However, traditional centralized training requires uploading raw Received Signal Strength Indicator (RSSI) data to a central server, which poses significant challenges regarding uplink communication bandwidth, energy consumption, and user privacy.

Federated Learning (FL) addresses privacy concerns by keeping data local, but standard aggregation methods like FedAvg still require clients to transmit full-precision (32-bit) model updates. In bandwidth-constrained wireless environments, this uplink communication cost becomes non-negligible, especially as the number of clients and communication rounds increases. Furthermore, Wi-Fi fingerprint data is often highly heterogeneous (Non-IID), where different clients (e.g., in different buildings or floors) possess distinct data distributions. This "Strong Non-IID" setting can cause client drift, leading to inconsistent local update directions that destabilize aggregation strategies relying on low-bit representations, such as binary voting.

This paper addresses the dual challenge of reducing uplink communication costs while maintaining localization accuracy under Strong Non-IID data conditions for Building-Floor dual-task Wi-Fi fingerprint localization.

2. Methodology

2.1 System Model and Dual-Task Formulation

The authors formulate the problem as a Building-Floor dual-task federated learning task.

Model Architecture: A lightweight dual-task Multi-Layer Perceptron (MLP) is employed. It consists of a shared feature extractor $g(\cdot)$ that processes the 520-dimensional RSSI vector, followed by two task-specific heads: a Building classifier $f_b(\cdot)$ and a Floor classifier $f_f(\cdot)$ .
Loss Function: The local loss combines Building and Floor classification losses, weighted by a hyperparameter $\alpha$ (set to 0.7 in experiments) to balance the tasks, as Floor prediction is more critical for the overall "Dual Accuracy."
Evaluation Metric: Dual Accuracy is the primary metric, defined as the proportion of test samples where both the Building and Floor labels are predicted correctly simultaneously.

2.2 Communication-Efficient Aggregation: Binary FedVote

To drastically reduce uplink traffic, the paper adopts a Binary FedVote-style aggregation strategy inspired by sign-based optimization.

Mechanism: Instead of uploading full-precision model updates ( $\Delta_k^t$ ), clients upload only the 1-bit sign direction ( $sign(\Delta_k^t)$ ) of their local updates.
Server Aggregation: The server computes a global voting direction by performing a weighted sum of the received binary signs and applying the sign function again. The global model is then updated using this voted direction and a predefined server step size.
Compression: This reduces the uplink communication from 32 bits per parameter to 1 bit per parameter, theoretically achieving a 32 $\times$ compression ratio compared to FedAvg.

2.3 Robustness Enhancement: FedProx Regularization

The authors identify that Binary FedVote-style aggregation is unstable under Strong Non-IID conditions because heterogeneous local data distributions cause conflicting update directions. To mitigate this without increasing communication costs, they integrate FedProx.

Local Objective: A proximal regularization term $\frac{\mu}{2} \|\theta - \theta_t\|^2$ is added to the local training objective.
Function: This term penalizes local models that deviate too far from the current global model $\theta_t$ . By restricting client drift, the local update directions become more consistent across heterogeneous clients, thereby stabilizing the 1-bit voting process on the server side.
Cost Implication: Since the uploaded information remains 1-bit binary data, the communication cost does not increase despite the added computational regularization.

3. Key Contributions

The paper outlines three primary contributions:

Dual-Task Formulation: The problem is framed as a Building-Floor dual-task FL problem, utilizing Dual Accuracy to evaluate the simultaneous correctness of both labels, which is more rigorous than evaluating tasks independently.
Binary FedVote Application: The application of 1-bit direction voting to federated Wi-Fi fingerprint localization. Under Moderate Non-IID settings, this approach reduces uplink communication from ~200.33 MB to ~6.26 MB while maintaining accuracy comparable to FedAvg.
FedProx-Enhanced Stability: The design of a FedProx-Binary FedVote mechanism. This method improves the robustness of low-bit aggregation under Strong Non-IID distributions by reducing client drift, thereby stabilizing the voting process without increasing uplink bandwidth.

4. Experimental Results

Experiments were conducted on the UJIIndoorLoc dataset with 10 clients, simulating both Moderate and Strong Non-IID data partitions.

4.1 Moderate Non-IID Setting

Performance: Binary FedVote-style achieved a Dual Accuracy of 0.9118, comparable to FedAvg (0.9100).
Communication: Uplink traffic was reduced from 200.33 MB (FedAvg) to 6.26 MB (Binary FedVote), a 32 $\times$ compression.
Observation: Confusion matrix analysis showed that the 1-bit voting strategy preserved the discriminative ability for Floor classification, with errors primarily occurring between adjacent floors, similar to the full-precision baseline.

4.2 Strong Non-IID Setting

Baseline Performance: Under Strong Non-IID, standard FedAvg dropped to a mean Dual Accuracy of 0.6695. The original Binary FedVote-style performed slightly better at 0.6976 but remained sensitive to data heterogeneity.
FedProx-Binary FedVote Performance: Introducing the FedProx term significantly improved stability, raising the mean Final Dual Accuracy to 0.8691 (with a standard deviation of 0.0358).
Efficiency: This improvement was achieved without increasing the uplink communication cost, which remained at 6.26 MB (1-bit per parameter).

5. Significance and Claims

The paper claims that the proposed FedProx-Binary FedVote method offers a practical trade-off between localization performance and uplink communication efficiency.

Communication Efficiency: It demonstrates that 1-bit direction voting is sufficient for dual-task Wi-Fi fingerprint localization under moderate heterogeneity, reducing bandwidth requirements by two orders of magnitude compared to full-precision methods.
Robustness: It addresses a critical gap in existing literature by showing that low-bit voting can be stabilized under Strong Non-IID conditions through local regularization (FedProx), rather than requiring complex server-side corrections or increased communication.
Applicability: The method is particularly relevant for wireless indoor localization scenarios where clients are bandwidth-constrained and data distributions are highly heterogeneous (e.g., different buildings or floors).

The authors conclude that while the current framework focuses on Building and Floor classification, future work could extend this to coordinate regression or more complex heterogeneous partitions, but the current study establishes a foundation for communication-efficient, robust federated indoor localization.