⚡ electrical engineering

MU-SHOT-Fi: Self-Supervised Multi-User Wi-Fi Sensing with Source-free Unsupervised Domain Adaptation

This paper proposes MU-SHOT-Fi, a source-free unsupervised domain adaptation framework that leverages permutation-invariant set prediction, occupancy-weighted information maximization, and spatial self-supervision to enable robust multi-user Wi-Fi activity recognition across diverse environments without requiring labeled target data.

Original authors: Ahmed Y. Radwan, Hina Tabassum

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmed Y. Radwan, Hina Tabassum

Original paper licensed under CC BY 4.0 (http://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

Imagine you are trying to teach a security guard (an AI model) to recognize what people are doing in a room just by listening to the Wi-Fi signals bouncing off the walls. This is called Wi-Fi sensing.

The problem is that the guard you trained in your living room (the Source) gets completely confused when you move them to an office (the Target). The furniture is different, the walls are different, and the Wi-Fi signals behave differently.

Even worse, in a real office, there isn't just one person moving around. There are multiple people doing different things at the same time. Their movements mix together in the Wi-Fi signal, creating a tangled mess. It's like trying to hear one person's voice in a crowded, noisy room where everyone is talking over each other.

This paper introduces a new system called MU-SHOT-Fi to solve two big problems:

The "New Room" Problem: How to make the AI work in a new environment without retraining it with new labeled data (which is often impossible due to privacy).
The "Crowded Room" Problem: How to figure out what each person is doing when there are multiple people, and how to stop the AI from just guessing "nobody is here" because that's the easiest answer.

Here is how it works, using simple analogies:

1. The "No-Label" Challenge (Source-Free Adaptation)

Usually, to teach an AI about a new room, you need a teacher to point at people and say, "That's walking, that's jumping." But in real life, you can't always have a teacher watching and labeling everything (privacy issues).

MU-SHOT-Fi is like a smart student who has already studied hard in the living room. When they walk into the office, they don't ask for a teacher. Instead, they look at the Wi-Fi signals and say, "I'm pretty sure this signal looks like walking, and that one looks like jumping, even though I've never seen this specific office before." They learn by trusting their own best guesses and refining them, a process called Self-Supervised Learning.

2. The "Tangled Signal" Problem (Multi-User Sensing)

In a single-person scenario, the AI just needs to guess one activity. But in a multi-user scenario, imagine the AI has 6 empty chairs (slots) to fill. It needs to decide who is sitting where and what they are doing.

The Trap: In most offices, most of the time, chairs are empty. If the AI just tries to be "diverse" (guessing all different activities), it might get confused and start guessing "nobody" for everyone because "nobody" is the most common answer. This is called model collapse.
The Solution (Occupancy-Weighted Maximization): The authors invented a special rule. They tell the AI: "Don't worry about the empty chairs. Only try to be diverse and confident about the chairs that actually have people in them."
- Analogy: Imagine a teacher grading a test. If a student leaves a question blank, the teacher ignores it. But if the student writes an answer, the teacher checks if it's unique and correct. MU-SHOT-Fi ignores the "empty chair" answers so the AI doesn't get lazy and just guess "empty" for everything.

3. The "Magic Mirror" (Rotation Self-Supervision)

To help the AI understand the new room without a teacher, the system uses a trick called Rotation Self-Supervision.

Analogy: Imagine you are looking at a picture of a person walking. If you turn the picture upside down or sideways, it's still the same person walking, just oriented differently.
The AI is trained to recognize that a Wi-Fi signal pattern is the same "activity" even if the signal is mathematically rotated (like turning a map). This forces the AI to learn the essence of the movement (the shape of the signal) rather than memorizing the specific noise of the old room. This helps it adapt to the new room.

4. Single Person vs. Multiple People

The paper also created a simpler version called SU-SHOT-Fi for when there is only one person.

In the single-person world, the AI can use a different trick called Contrastive Predictive Coding (CPC). This is like listening to a song and predicting the next note based on the previous ones. It helps the AI understand the rhythm of a single person's movement over time.
The Twist: The authors found that this "rhythm" trick does not work for multiple people. Why? Because when two people move, their rhythms get mixed up. Trying to predict the "next note" of a tangled mess confuses the AI. So, for crowded rooms, they stick to the "Magic Mirror" trick and the "Occupancy" rule.

The Results

The researchers tested this on real data:

Multi-User (WiMANS dataset): When they moved the AI from a classroom to a meeting room, or changed the Wi-Fi frequency (from 2.4GHz to 5GHz), the old AI failed completely (0% accuracy on exact matches). MU-SHOT-Fi recovered and got the job done, correctly identifying who was doing what in the new room.
Single-User (Widar 3.0 dataset): The single-person version also improved, getting better at recognizing gestures like "clapping" or "drawing a zigzag" in new rooms.

Summary

MU-SHOT-Fi is a smart system that allows Wi-Fi sensors to adapt to new, crowded environments without needing a human to label every single moment. It does this by:

Ignoring empty spaces so it doesn't get lazy.
Learning to recognize movements even when the signal is "rotated" or changed by the environment.
Knowing that "rhythm" tricks work for solo acts but fail in a crowd.

It's essentially teaching a Wi-Fi sensor to be a flexible, self-taught detective that can solve mysteries in a new, crowded room without ever being told the answers.

Technical Summary: MU-SHOT-Fi: Self-Supervised Multi-User Wi-Fi Sensing with Source-free Unsupervised Domain Adaptation

Problem Statement
Deep learning (DL) has advanced Wi-Fi Channel State Information (CSI)-based Human Activity Recognition (HAR) by learning spatio-temporal features without manual engineering. However, DL models suffer from poor generalization across environments (domain shifts) caused by changes in physical layout, user characteristics, hardware, and frequency bands. This challenge is exacerbated in multi-user settings where concurrent activities create "signal entanglement," making it difficult to attribute CSI variations to specific individuals. Furthermore, practical deployments face privacy constraints that prevent access to labeled source data during target domain adaptation. While Unsupervised Domain Adaptation (UDA) exists, it requires labeled source data. Source-Free Unsupervised Domain Adaptation (SFUDA) addresses this but faces three critical hurdles in multi-user Wi-Fi sensing:

Class Imbalance: "No person" (empty) slots dominate datasets, causing standard information maximization to collapse into predicting the dominant empty class.
Permutation Invariance: Multi-user states are unordered sets (e.g., [walk, sit] is equivalent to [sit, walk]), rendering standard clustering-based pseudo-labeling and slot-level matching infeasible without ground truth.
Signal Entanglement: Multi-user CSI reflects superposed activities, violating the stable feature-to-label mappings assumed by many adaptation techniques.

Methodology
The authors propose MU-SHOT-Fi, a source-free framework designed for both single-user and multi-user scenarios. The framework operates in two stages: source training and target adaptation.

Source Training: The model is trained on labeled source data using a permutation-invariant set prediction approach. Instead of assigning activities to fixed slots, the model predicts a set of activities across $M$ slots. A Hungarian matching algorithm optimally aligns predicted slots with ground-truth labels to compute the cross-entropy loss, handling variable occupancy and unordered user states.
Target Adaptation (SFUDA): In the target domain, only unlabeled data is available. The classifier is frozen, and the backbone is adapted using two self-supervised objectives:
1. Occupancy-Weighted Information Maximization: To prevent collapse into the dominant "no person" class, the authors introduce a modified information maximization loss. Unlike standard methods that enforce diversity across all slots, this approach weights diversity regularization by the estimated occupancy probability of each slot. It excludes the "no person" class from the marginal entropy term, ensuring the model focuses on diversifying predictions only for likely-occupied slots.
2. Rotation-Based Spatial Self-Supervision: To learn domain-invariant features, the framework employs a binary rotation prediction task (0° vs. 180°) on the CSI frequency-time grid. This exploits the structural properties of CSI signals to align source and target representations without labels.
Single-User Variant (SU-SHOT-Fi): For single-user scenarios, the framework is customized by replacing occupancy-weighted maximization with standard information maximization, enabling k-nearest centroid pseudo-labeling (which is stable due to balanced class distributions), and adding Contrastive Predictive Coding (CPC) to exploit temporal consistency in CSI sequences.

Key Contributions

MU-SHOT-Fi Framework: The first SFUDA framework specifically designed for multi-user Wi-Fi sensing, agnostic to CSI preprocessing (amplitude or phase ratio).
Occupancy-Weighted Information Maximization: A novel loss function that mitigates class imbalance by dynamically weighting diversity regularization based on slot occupancy, preventing model collapse toward the "no person" class.
Permutation-Invariant Learning: The integration of Hungarian matching during source training and its adaptation to the target domain without explicit slot correspondence, addressing the unordered nature of multi-user states.
Spatial Self-Supervision: The use of rotation prediction on CSI frequency-time structures to learn domain-invariant features without labels.
Comprehensive Evaluation: Extensive experiments on the WiMANS (multi-user) and Widar 3.0 (single-user) datasets across cross-environment, cross-frequency, and cross-orientation shifts.

Experimental Results

Multi-User (WiMANS): Under combined domain shifts (cross-environment and cross-frequency), the source-only model achieved only 19.61% slot-wise accuracy. MU-SHOT-Fi improved this to 41.97%, while reducing Occupancy Mean Absolute Error (MAE) from 3.00 to 2.12. It also achieved a 2.88% Exact Match accuracy (vs. 0.71% for baselines), representing a massive improvement over random chance in the combinatorial output space.
Single-User (Widar 3.0): SU-SHOT-Fi improved average accuracy from 80.18% (source-only) to 85.75% across various shifts, outperforming standard SFUDA baselines (SHOT-IM and SHOT++).
Ablation Studies:
- Adding CPC to the multi-user model degraded performance, as temporal consistency conflicts with the permutation-invariant slot reordering required by Hungarian matching.
- A pooled parallel classifier architecture outperformed a query-based transformer decoder (DETR-style) in adaptation, as the latter tended to exploit the "no person" majority class rather than learning activity features.

Significance and Claims
The paper claims that MU-SHOT-Fi effectively recovers multi-user exact-activity classification performance under large domain shifts while maintaining accurate occupancy estimation. The authors emphasize that their approach prevents the model collapse toward dominant classes—a common failure mode in multi-user SFUDA—by explicitly modeling the occupancy structure of the output space. The work highlights that while standard adaptation techniques (like adversarial learning or simple entropy minimization) fail in multi-user settings due to signal entanglement and class imbalance, the proposed occupancy-weighted diversity and permutation-invariant learning enable stable adaptation. The source code is publicly available to facilitate further research in privacy-preserving, multi-user Wi-Fi sensing.