⚡ electrical engineering

TinyML for On-Device and Edge Analytics in Wireless Networks: A Survey of Deployments, Opportunities, and Concept-Drift Mitigation

This paper surveys the deployment opportunities and frameworks for tinyML in wireless networks and proposes a federated learning-based approach with specialized mechanisms to mitigate concept drift and support model updates on both battery-powered and batteryless edge devices.

Original authors: Prasoon Raghuwanshi, Vimal Bhatia, Sridhar Iyer, Matti Latva-aho, Onel Luis Alcaraz López

Published 2026-07-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Prasoon Raghuwanshi, Vimal Bhatia, Sridhar Iyer, Matti Latva-aho, Onel Luis Alcaraz López

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 the world of wireless networks (like Wi-Fi, 5G, and future 6G) as a massive, bustling city. Currently, most of the "thinking" in this city happens in giant, centralized skyscrapers called the Cloud. Every time your smartwatch, car, or sensor needs to make a decision, it sends a message up to the skyscraper, waits for an answer, and then acts.

This paper argues that it's time to build tiny, intelligent brains right inside the devices themselves (the "end-devices" and "edge nodes"). This technology is called TinyML.

Here is a simple breakdown of what the paper says, using everyday analogies:

1. What is TinyML? (The "Pocket-Sized Chef")

Think of a standard Machine Learning model (like the ones running in the cloud) as a giant, 5-star restaurant kitchen. It has massive ovens, huge fridges, and a team of 50 chefs. It can cook anything, but it takes up a whole building and costs a fortune to run.

TinyML is like a master chef with a portable camping stove.

The Problem: You can't fit the giant kitchen into a tiny backpack (your phone or sensor).
The Solution: TinyML takes that giant recipe and shrinks it down. It uses tricks like pruning (cutting out unnecessary ingredients), quantization (measuring ingredients in smaller, simpler units), and distillation (teaching a small student chef to mimic the big master chef).
The Result: The device can now "cook" (make decisions) right there in the field, without needing to call the skyscraper. This saves energy, protects privacy (data stays local), and works even if the internet goes down.

2. Where is it being used? (The "Already Open Shops")

The paper reviews where this "portable chef" is already working in wireless networks:

Smart Filters: Instead of sending every photo a camera takes to the cloud, the device checks the photo first. If it's just a blurry wall, it deletes it. If it sees a person, it sends the photo. This saves bandwidth.
Sensor Fixers: Sensors (like temperature gauges) often get "grumpy" and give weird readings as they age. TinyML acts like a personal trainer for the sensor, constantly correcting its readings so they stay accurate without needing a human to recalibrate them.
Radar Eyes: Self-driving cars use radar. Usually, the raw radar data is too heavy to process on the car's computer. TinyML allows the car to process this data instantly to spot pedestrians or other cars, even in the rain.
Security Guards: TinyML can sit on a network router and act like a bouncer, spotting suspicious traffic patterns (intruders) immediately and blocking them before they reach the main server.

3. What's New? (The "Unexplored Territories")

The paper suggests many new places where these tiny brains could work, which haven't been fully explored yet:

Wireless Power Charging: Imagine charging a drone or a medical implant wirelessly. TinyML could act as a smart conductor, adjusting the power beam in real-time to ensure the drone gets charged efficiently, even if it moves around or if there are obstacles.
Goal-Oriented Chat: Instead of a device sending everything it sees, TinyML helps it decide what to say. If a sensor is watching a machine, it only sends a message if the machine starts acting weird (an anomaly), rather than sending a constant stream of "everything is fine" data.
Underwater Communication: Light behaves strangely underwater. TinyML can act as a translator, figuring out how to send messages clearly through murky water by predicting how the light will bend and scatter.

4. The Big Challenge: "Concept Drift" (The "Changing Weather")

Here is the tricky part. Imagine you trained your portable chef to recognize summer fruits. But then, the season changes to winter, and the fruits look different. The chef is now confused and makes mistakes. In data science, this is called Concept Drift. The world changes, but the model stays the same.

The paper proposes two ways to fix this:

A. For Battery-Powered Devices (The "Study Group"):

Federated Learning: Instead of sending all your data to the cloud to retrain the model, the device learns from its own local data. It then sends just the "lessons learned" (mathematical updates) to a central hub. The hub combines lessons from thousands of devices to create a smarter "Global Model" and sends it back.
The "Tiny" Twist: Since these devices have very little memory, the paper suggests special tricks to make this training possible on tiny chips, like Gradient Scaling (adjusting the math so small numbers don't get lost) and Sparse Updates (only updating the most important parts of the brain, not the whole thing).

B. For Batteryless Devices (The "Intermittent Workers"):
Some devices run on energy harvesting (like solar power or vibration). They might have power for 10 seconds, then run out for an hour.

The Problem: If a power outage happens while the device is updating its brain, the update might get corrupted, and the device could "die."
The Solution: The paper suggests a Fault-Tolerant Bootloader. Think of this as a safety net.
- Checkpointing: The device saves its progress frequently, like a video game save point. If the power cuts out, it can restart from the last save, not from the beginning.
- Intermittent-Aware Updates: The device knows how to update its code in small, safe chunks, ensuring that even if the power flickers, the new code doesn't get scrambled.

Summary

This paper is a roadmap for the future of wireless networks. It argues that we shouldn't just rely on giant cloud servers. Instead, we should empower our devices with TinyML—small, efficient brains that can think for themselves.

Current State: TinyML is already helping with sensors, radar, and security.
Future Potential: It could revolutionize wireless charging, underwater comms, and smart resource management.
The Hurdle: The world changes (Concept Drift), so these tiny brains need to learn continuously.
The Fix: We have new methods to let them learn locally (Federated Learning) and survive power cuts (Fault-Tolerant Bootloaders), ensuring they stay smart and reliable forever.

Technical Summary: TinyML for On-Device and Edge Analytics in Wireless Networks

1. Problem Statement

The proliferation of microcontroller units (MCUs) in wireless networks and Internet of Things (IoT) devices necessitates a shift toward ubiquitous intelligence to enable real-time, adaptive, and autonomous operations. However, deploying machine learning (ML) on these resource-constrained end-devices and edge nodes presents significant challenges:

Resource Constraints: End-devices often have limited memory (kilobytes of RAM), processing power, and energy (battery-powered or energy-harvesting batteryless).
Dynamic Environments: Wireless networks operate under dynamic conditions where data distributions change over time, leading to concept drift. This renders pre-trained models ineffective without retraining.
Energy and Connectivity: Continuous connectivity to centralized clouds for data analysis is energy-intensive and introduces latency. Furthermore, batteryless devices face intermittent power availability, complicating model updates and state preservation.
Gap in Literature: Existing surveys focus heavily on model compression, hardware platforms, or specific application domains (e.g., healthcare, computer vision). There is a lack of comprehensive analysis regarding the deployment possibilities of TinyML specifically within wireless network problems, particularly concerning the integration with batteryless devices and the mitigation of concept drift via Federated Learning (FL).

2. Methodology

The paper employs a systematic literature review methodology to identify, analyze, and synthesize existing research and propose new deployment avenues.

Search Criteria: The authors searched databases (IEEE Xplore, Elsevier, Springer, etc.) using keywords including "wireless networks," "TinyML," "MCU," "on-device inference," and "batteryless end-device."
Inclusion/Exclusion: Empirical studies implementing TinyML on MCU-based devices within wireless contexts were included. Cloud-only implementations were excluded.
Scope: The review covers 146 candidate studies, categorized into:
1. State-of-the-art survey papers (18).
2. TinyML frameworks for low-power devices (42).
3. Tapped and untapped deployment possibilities in wireless networks (63).
4. Concept drift mitigation techniques (12).
Analysis Framework: The paper categorizes TinyML algorithms by target hardware (edge nodes, battery-powered devices, batteryless devices) and analyzes their suitability for specific wireless tasks. It also proposes technical architectures for model updates, including Federated Learning (FL) and specific techniques for intermittent computing.

3. Key Contributions

A. Survey of TinyML Algorithms and Frameworks

The paper categorizes existing TinyML solutions based on the target hardware constraints:

Low-Power Edge Nodes: Utilize frameworks like OnceNAS and Weight Factorization for tasks such as cooperative spectrum sensing, automatic modulation classification (AMC), and wireless channel estimation. Reinforcement Learning (RL) frameworks like PoPS are identified for IoT device scheduling.
Battery-Powered End-Devices: Highlight algorithms such as Memory-aware Hybrid, TinyNS, and TyCNN for anomaly detection and semantic encoding. Classical ML methods (Decision Trees, Random Forests) are noted for their computational efficiency on MCUs.
Batteryless End-Devices: Due to severe energy constraints, the paper suggests using single-class classification (via class-specific models) and shifting complex RL training to edge nodes, with only actions sent to the device. Frameworks like LBPNet (using bit-shifting instead of convolutions) and Genesis (minimization tool) are cited for viability.

B. Tapped Deployment Possibilities (Existing Literature)

The paper reviews documented applications of TinyML in wireless networks:

Semantic Data Transmission: Using TinyDL (e.g., EtinyNet) to filter images based on semantic queries before transmission, reducing bandwidth usage.
Sensor Linearization: Correcting non-linear sensor responses (e.g., air/light pollution) on-device to improve data quality and reduce retransmissions.
Radar Processing: Using compressed CNNs for object detection in automotive radar, preserving information lost in traditional point-cloud processing.
RF Fingerprinting & Localization: Enabling edge-based positioning and indoor localization (using Two-Way Ranging) with reduced latency and enhanced privacy.
Signal Processing: Implementing TinyDL-based pre-distorters for MIMO-OFDM transmitters and CSI-free data symbol detection for energy-scarce receivers.
Security: Embedding Network Intrusion Detection Systems (NIDS) on devices to detect signature-based and anomaly-based attacks locally.

C. Untapped Deployment Possibilities

The authors identify several under-explored domains where TinyML could offer significant benefits:

Near-Field Wireless Power Transfer (WPT): Using TinyDL to estimate system characteristics (load resistance, coupling coefficient) for sensorless receivers, managing TX-coil activation in multi-TX systems, and detecting unintended loads (ULD) in Unmanned Aircraft Systems (UAS).
Backscatter-Assisted WSNs: Using TinyRL to optimize reflection coefficients for spectral efficiency.
Goal-Oriented Communication (GoC): Applying TinyRL for device scheduling in dynamic monitoring and control systems to minimize error in query responses or actuation.
Local Wake-Up Strategies: Using TinyRL to manage duty cycles for batteryless devices to minimize missed alarm events.
Underwater Visible Light Communication (UVLC): Employing TinyDL for channel estimation to counteract noise and distortion.
Cognitive Radio Networks: Utilizing TinyDL for Cooperative Spectrum Sensing (CSS) and TinyRL for distributed resource allocation in Device-to-Device (D2D) networks.

D. Mitigation of Concept Drift via Federated Learning

The paper proposes a Federated Learning (FL) framework to update TinyML models without centralized data collection, addressing concept drift:

Battery-Powered Devices:
- Centralized TinyFL: A central entity distributes a global model; devices perform on-device training and send trainable parameters back.
- Backpropagation Revamping: To overcome memory constraints during training, the paper details three techniques:
  1. Gradient Scaling: Rescaling gradients to stabilize optimization in quantized (int8) models.
  2. Sparse Update: Updating only a subset of parameters (e.g., biases or final layers) to reduce memory footprint.
  3. Tiny Training Engine: A compilation-based approach that generates a static backward graph, prunes redundant operators, and reorders execution to minimize peak memory and latency.
Batteryless Devices:
- Over-the-Air Aggregation: Utilizing waveform superposition to aggregate model parameters during transmission.
- Intermittent-Aware Modify Operation: A mechanism to handle code updates in stages if the new code block is larger than the replacement block, preventing code fragmentation.
- Update-Aware Checkpointing: Dividing SRAM into stack and update buffer regions to save state during power interruptions.
- Fault-Tolerant Bootloader: A bootloader that supports checkpointing and no-reboot operations, significantly reducing energy consumption during model updates compared to baseline approaches (e.g., FLoRa).

4. Results and Findings

Performance of Revamped Training:
- Gradient Scaling: Restores the accuracy of int8 quantized models to be comparable to fp32 models (e.g., MCUNet-5FPS accuracy on Flowers dataset: 84.5% for int8 vs. 89.1% for int8+scaling).
- Sparse Update: Reduces extra memory footprint by ~~3.6x compared to updating the last $k$ layers while maintaining high accuracy (~~72% vs. ~70%).
- Tiny Training Engine: Reduces training latency by ~23x and peak memory by ~21x compared to full backpropagation (e.g., MobileNetV2-w0.35: 373ms latency and 141kB peak memory vs. 8501ms and 2939kB).
Batteryless Updates: The proposed framework using intermittent-aware modify operations and fault-tolerant bootloaders reduces the energy required to restart a model update task after a power interruption from ~~67.5mJ (FLoRa baseline) to ~2.5mJ. It also enables finite total update times (~~1350ms) where baselines fail.
Deployment Viability: The paper demonstrates that TinyML can achieve high accuracy in wireless tasks (e.g., 99.5% accuracy for indoor presence detection, ~88% for RF fingerprinting positioning) while operating within strict memory (kB) and energy constraints.

5. Significance and Claims

The paper positions itself as a timely and essential survey to bridge gaps in the rapidly emerging interdisciplinary domain of TinyML in wireless networks. Its significance lies in:

Comprehensive Scope: Unlike previous surveys, it explicitly addresses the integration of TinyML with batteryless devices and low-power edge nodes, which are critical for future 6G and sustainable IoT.
Concept Drift Solution: It provides a concrete technical roadmap for handling concept drift in resource-constrained environments using FL and specialized update mechanisms (checkpointing, fault-tolerant bootloaders).
Untapped Potential: It identifies and details specific, previously overlooked application domains (e.g., WPT optimization, GoC, UVLC) where TinyML can solve critical wireless communication problems.
Practical Implementation: By detailing the "how-to" of on-device training (gradient scaling, sparse updates) and intermittent computing, the paper moves beyond theoretical discussion to provide actionable engineering strategies for deploying adaptive intelligence on the edge.

The authors conclude that while TinyML is poised to become ubiquitous, realizing its full potential in 6G networks requires addressing protocol-level challenges related to model integration, dataset availability, energy profiling, and trustworthy decision-making.