AI That Detects Enemy Signals in Real Time
Here’s How It Works
Most people think of artificial intelligence as something that recommends movies or writes emails. I wanted to push it somewhere more challenging — into the electromagnetic spectrum, where signals are invisible, enemies are adaptive, and milliseconds matter.
This is the story of how I designed a deep learning system that listens to the air, identifies threats it has never seen before, and responds autonomously.
The problem I set out to solve
Modern electronic warfare isn’t just about brute-force jamming. It’s about recognition. A system that can’t tell a commercial drone from a military one, or a known signal from an unknown one, is flying blind.
Traditional signal classification relies on hand-crafted features — engineers manually defining what a “threat” looks like. That approach fails the moment an adversary uses a signal you’ve never catalogued.
My question was simple: what if the system could learn?
The architecture: two models, one brain
I built the system around two complementary deep learning models that work in tandem:
1D Convolutional Neural Network (1D-CNN) — This handles known threats. Trained on labelled RF signal datasets (commercial drones, specific transmitters), it classifies incoming signals with high confidence and low latency. Think of it as the experienced analyst who instantly recognizes familiar patterns.
Autoencoder — This handles the unknown. Rather than classifying, it learns what “normal” looks like and raises an alert when something deviates from that norm. It’s the equivalent of a sensor that says: “I don’t know what this is — but I know it shouldn’t be here.”
The fusion of these two outputs creates a decision layer. If the signal is known → classify and respond. If the signal is anomalous → flag for autonomous countermeasure.
Software-defined radio: the ears of the system
Raw RF data enters the system via a Software-Defined Radio (SDR) device. Unlike traditional radio hardware that’s locked to specific frequencies and protocols, an SDR is programmable — it captures the raw electromagnetic environment and passes it directly to software.
This is where the pipeline starts: SDR captures raw signal → preprocessing converts it into model-readable format → the dual-model system takes over. The entire flow is designed to operate in real time.
When a threat is confirmed: autonomous countermeasures
Here’s where the project steps beyond classification. Once a threat signal is confirmed, the system doesn’t just log it — it responds.
Get Fırat Yağmur’s stories in your inbox
Join Medium for free to get updates from this writer.
In a simulation environment, it initiates one of two Electronic Countermeasure (ECM) scenarios:
Jamming
Broadcast interfering signals to disrupt the threat’s communication link
Spoofing
Transmit a fake signal that mimics a legitimate one — deceiving the threat system
The entire cycle — detect, classify, decide, respond — is autonomous. No human in the loop for routine threats.
Why this matters beyond the lab
We are living through a period where unmanned aerial vehicles, rogue transmitters, and electronic warfare tools are increasingly accessible. The gap between detecting a threat and responding to it can be the difference between mission success and catastrophic failure.
Systems that rely purely on pattern databases will always be a step behind an adaptive adversary. Systems that can detect anomalies — signals they’ve never seen — change that equation.
This project is a proof of concept that deep learning and SDR technology, combined thoughtfully, can form the foundation of next-generation autonomous electronic defense.
What I learned building this
Three things stand out:
Signal preprocessing is everything. Raw RF data is messy. Getting it into a format a neural network can actually learn from took more iteration than the model training itself.
Autoencoders are underrated. Most engineers reach for classifiers. But for anomaly detection in open-world environments, the autoencoder approach — learning normality rather than labelling threats — is far more robust.
Fusion beats single models. Neither the CNN nor the autoencoder alone was sufficient. The real power came from combining their outputs into a unified decision layer — a pattern I’ll carry into every multi-modal system I build.
What’s next
The current system operates in simulation. The next phase involves hardware-in-the-loop testing — deploying the models on edge hardware connected to a real SDR, and validating the countermeasure responses against live (controlled) RF environments.
I’m also exploring how transformer-based architectures — which have reshaped NLP — could be applied to temporal RF signal sequences. The parallels between language tokens and signal samples are compelling.
