Inspiration
In complex RF environments, friendly, hostile, and civilian signals coexist in the same spectrum. However, real-world systems often only have labeled data for friendly signals. We built FindMyForce to solve this problem by creating a machine learning system that can recognize friendly signals while detecting unknown emitters that may be hostile or civilian.
What it does
In complex RF environments, friendly, hostile, and civilian signals coexist in the same spectrum. However, real-world systems often only have labeled data for friendly signals. We built FindMyForce to solve this problem by creating a machine learning system that can recognize friendly signals while detecting unknown emitters that may be hostile or civilian.
How we built it
FindMyForce analyzes RF IQ signal data and classifies signals into friendly types while detecting unknown transmissions. If the model is confident, it labels the signal as one of the known friendly emitters (Radar-Altimeter, Satcom, or Short-Range). If confidence is low, the system flags the signal as unknown and analyzes its characteristics to determine whether it is likely hostile radar, a jammer, or civilian radio.
Challenges we ran into
We extracted around 90 signal features from raw IQ data, including amplitude statistics, phase patterns, spectral energy, and timing characteristics. These features are fed into a HistGradientBoostingClassifier, which learns patterns of friendly signals. During prediction, the model outputs probabilities for each class. A confidence threshold is used for open-set detection: high confidence indicates a friendly signal, while low confidence triggers further analysis to identify potential hostile or civilian emitters.
Accomplishments that we're proud of
The biggest challenge was that hostile and civilian signals were not included in the training data. We solved this by implementing confidence-based novelty detection, allowing the system to recognize when a signal does not match any known friendly pattern.
What we learned
We built a system capable of open-set RF classification, meaning it can recognize known signals while detecting previously unseen ones. The model combines machine learning with signal-processing techniques to operate in noisy RF environments.
What's next for FindMyForce
Next, we want to improve detection using deep learning on spectrograms, add stronger anomaly detection models, and support real-time RF monitoring for dynamic environments.
Log in or sign up for Devpost to join the conversation.