How Can AI Stop Jamming Attacks in Power Line Networks?

Power‑Line Communication (PLC) lets data travel through existing electricity cables, saving the cost of building new networks. However, PLC lines are often left unattended, making them vulnerable to malicious jamming signals that can disrupt communication. Researchers have now designed an AI‑based system that spots these attacks before they degrade the link. Two scenarios are considered. In the first, the receiver knows the jammer’s impact: a Convolutional Neural Network (CNN) classifier, trained on both jammed and clean signals, detects jamming with over 99% accuracy, even when the jammer is 68 meters away. In the second scenario, the receiver lacks prior knowledge: a one‑class auto‑encoder learns to reconstruct only normal traffic, flagging any deviation as an anomaly, i.e., a possible jammer. The key advantage is that the system trains solely on regular PLC traffic, without needing simulated attacks. This enables utilities to monitor their lines in real time, quickly locate the source of interference, and act before the grid suffers. A clear case of turning a weakness into a security strength.

PLC networks operate on channels characterized by impulsive noise, frequency‑dependent attenuation, and reflections caused by non‑linear loads. Jamming can take various forms: broadband signals that overwhelm the entire spectrum, or narrowband interference targeting pilot frequencies. To train the CNN, researchers gathered a dataset of 10,000 signal segments, each sampled at 1 MS/s, labeled as "jammed" or "clean." The network employs three convolutional blocks (3×3 kernels, stride 1) followed by max‑pooling and two fully‑connected layers, with ReLU activations and 30% dropout to prevent overfitting. The auto‑encoder consists of a three‑level encoder compressing the signal into a 64‑dimensional latent space, and a symmetric decoder that reconstructs the input. The anomaly threshold is set based on the reconstruction loss distribution on training data, chosen at the 95th percentile. Results show an ROC curve with AUC = 0.998 for the CNN and 0.985 for the auto‑encoder, confirming the ability to detect jammers even at 68 m distance, where the signal‑to‑noise ratio drops below –12 dB. Moreover, the model requires only 5 minutes of training on a consumer‑grade GPU, making it feasible for utilities. The approach scales: by adding new normal traffic segments, the system adapts to load variations and new network topologies, maintaining high accuracy without needing specific attack data.