⚡ electrical engineering

Characterizing and Mitigating the Effects of Device Temperature on RF Fingerprinting Accuracy

This paper proposes a novel temperature-aware Radio Frequency Fingerprinting framework that integrates device temperature data into the learning process to significantly improve authentication accuracy and robustness against temperature-induced signal variations, as validated on a real-world Bluetooth Low Energy dataset.

Original authors: Haytham Albousayri, Bechir Hamdaoui

Published 2026-07-29

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Haytham Albousayri, Bechir Hamdaoui

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 invisible air around us is a crowded dance floor where billions of tiny electronic devices are constantly whispering secrets to each other. To keep this party safe, we need a way to know exactly who is dancing. This is the world of Radio Frequency Fingerprinting (RFFP). Think of every device—your phone, a smartwatch, a Bluetooth speaker—as having a unique "voice" or "handshake" caused by tiny, accidental imperfections in its hardware. Just as no two human voices are exactly alike, no two electronic chips are perfect copies; they all have slight quirks that leave a distinct mark on the signals they send. Usually, security systems use these quirks to say, "Yes, that's definitely your phone," without needing a password. But there's a catch: these electronic voices can change their tone depending on how hot the device gets, much like a singer's voice might crack if they run a marathon in the summer heat. If the security system doesn't know the device is hot, it might mistake your phone for a stranger's, causing a security failure. This is the puzzle scientists are trying to solve: how do we keep the "voice" recognition working even when the device's temperature shifts?

Enter a team of researchers who decided to stop ignoring the heat. They discovered that existing methods for identifying devices were like trying to recognize a friend in a foggy room without knowing the weather; if the temperature changed, the recognition system got confused and failed. The researchers found that a key part of a device's signal, called the Carrier Frequency Offset (CFO), acts like a thermometer, shifting significantly as the device warms up. In their experiments with 12 Bluetooth Low Energy (BLE) devices, they saw that when a device heated up from a cool 30°C to a warm 56°C during a 20-minute session, the accuracy of standard identification systems could plummet from nearly perfect to a disastrous 40%.

To fix this, the team proposed a clever new framework called Temperature-Aware RFFP. Instead of just listening to the signal, their system also "asks" the device, "How hot are you right now?" They designed a special data packet where the device includes its own internal temperature reading alongside its usual signal. Then, they fed both the signal and the temperature number into a smart computer brain (a deep learning model) that learned to recognize the device while accounting for the heat. It's like teaching a bouncer to recognize a VIP not just by their face, but by knowing, "Oh, it's hot outside, so their face might look a bit flushed today, but it's still them."

The results were striking. When they tested their new system against older methods that ignored temperature or tried to mathematically remove the heat effects, the temperature-aware approach shined. In tests where the device was heated to unseen temperatures (up to 56°C) and placed in different locations (some with clear paths, some blocked by metal), the new system maintained an accuracy of over 97%. In contrast, the older "temperature-unaware" systems crashed to around 40% accuracy, and even the "temperature-invariant" systems (which tried to filter out the heat) dropped to about 90% or lower. The researchers suggest that by explicitly telling the computer the temperature, the model can learn the "trajectory" of how the device's voice changes as it heats up, rather than just memorizing a static snapshot. They even made their dataset of over 33,000 signal frames from 12 devices available to the public, hoping others will use this "hot" data to build even smarter, more reliable security systems for the future.

Technical Summary: Characterizing and Mitigating the Effects of Device Temperature on RF Fingerprinting Accuracy

Problem Statement
Radio Frequency Fingerprinting (RFFP) has emerged as a robust method for device authentication by exploiting hardware-specific impairments embedded in transmitted signals. While Deep Learning (DL) has enhanced RFFP's ability to generalize across varying channels, locations, and timeframes, existing methods largely overlook a critical degradation factor: device temperature. Internal and environmental temperature changes significantly alter hardware impairments, such as Carrier Frequency Offset (CFO), causing the device signatures to shift. Consequently, RFFP systems trained at specific temperatures often suffer severe performance drops when deployed under different thermal conditions. Previous studies have noted this sensitivity but have offered limited solutions, such as removing CFO features (which assumes temperature affects only CFO) or training separate models for discrete temperature points (which lacks scalability for continuous temperature variations).

Methodology
To address these limitations, the authors propose a novel Temperature-Aware RFFP framework. The core of this approach involves explicitly incorporating real-time device temperature information into the learning process as an additional input feature.

System Model: The framework treats the received signal $y_k[n]$ and the device temperature $T_k$ as a joint input vector $[y_k[n]; T_k]$ for the classifier. This allows the model to learn the trajectory of feature changes induced by temperature rather than relying on static representations.
Payload Design: To facilitate real-time temperature sharing, the authors designed a custom Bluetooth Low Energy (BLE) frame. The device's internal temperature is measured by its onboard sensor, encoded into the last two bytes of the payload field, and transmitted over-the-air (OTA) alongside the standard advertisement packets.
Deep Learning Architecture: The classification engine utilizes a Convolutional Neural Network (CNN) encoder to process the raw I/Q signal samples, generating feature embeddings. These embeddings are flattened and concatenated with the temperature measurement. The combined vector is then passed through a Multi-Layer Perceptron (MLP) to output the device identity probability distribution.
Experimental Setup: The framework was evaluated using a real-world dataset collected from 12 ESP32-C3 BLE devices and a USRP B210 receiver. Data was gathered across three distinct environments: Line-of-Sight (LoS) at 1.5m, Non-Line-of-Sight (NLoS) with a metal obstruction, and LoS at 7.5m in a rich fading environment. The dataset captures the device warm-up phase (where temperature rises from ~30°C to ~56°C) and subsequent steady-state operation.

Key Contributions
The paper outlines four primary contributions:

Scalable Framework: A temperature-aware RFFP system that accepts temperature as a continuous input feature, overcoming the limitations of discrete-temperature models and enabling robust operation across a wide thermal range.
Comprehensive Evaluation: Validation on a diverse, real-world dataset spanning multiple devices, locations (LoS and NLoS), and environmental conditions.
Performance Improvement: Experimental results demonstrating that the proposed method significantly outperforms existing baselines, particularly in unseen temperature and environmental conditions.
Open Data: The public release of the collected dataset to support reproducibility and benchmarking in the RFFP community.

Results
The study compares the proposed Temperature-Aware approach against two baselines:

Temperature-Unaware RFFP: Uses only the signal without temperature data.
Temperature-Invariant RFFP: Removes residual CFO and Phase Offset (PO) from the signal prior to classification.

Key findings include:

Accuracy Under Thermal Shift: When trained on data from the first 3 minutes (30°C–52°C) and tested on data from later times (up to 56°C), the Temperature-Unaware baseline's accuracy dropped to approximately 40%, and the Temperature-Invariant baseline dropped to roughly 90%. In contrast, the proposed Temperature-Aware framework maintained accuracy above 97%.
Cross-Location Generalization: The proposed framework demonstrated superior robustness when tested in locations different from the training location. While baselines suffered drastic accuracy declines (e.g., dropping to 20–30% in cross-location tests), the proposed method experienced only a modest decline (from ~97% to ~92%).
Model Efficiency: Among various DL architectures tested (CNN, Transformer, GRU, LSTM), the CNN architecture was selected for its balance of high accuracy and computational efficiency (lowest inference time).

Significance and Claims
The paper claims that ignoring temperature information leads to significant performance degradation in RFFP systems, especially under unseen conditions. By integrating temperature as an explicit feature, the proposed framework enables the classifier to effectively capture the underlying dynamics of hardware impairments rather than relying on static signal features. The authors assert that this approach provides a scalable solution for device authentication that remains robust across continuous temperature variations and diverse deployment scenarios, addressing a critical gap in current RFFP research. The work suggests that future RFFP systems must account for environmental and internal device variables to achieve reliable real-world deployment.

More like this