Abstract
Despite extensive research on jamming attacks in wireless communication systems, the potential of machine learning for amplifying these threats or mitigating them remains largely untapped. A primary obstacle has been the lack of a suitable framework that captures the complexities of modern standards. We introduce PyJama, a fully-differentiable open-source library that adds jamming and anti-jamming functionality to NVIDIA Sionna. By making the entire communication chain differentiable, PyJama enables the use of gradient-based optimization to develop sophisticated jamming attacks and robust defenses. We demonstrate its utility through realistic MIMO simulations involving OFDM waveforms, forward error correction, and mobility.
Introduction
Traditional research into wireless electronic warfare often relies on simplified models, such as basic signal-to-interference-plus-noise ratio (SINR) calculations. These "toy" models fail to account for the intricate structures of modern systems, including Orthogonal Frequency Division Multiplexing (OFDM) resource grids, pilot-based channel estimation, and the physical mobility of nodes. Furthermore, the absence of a unified, differentiable pipeline has forced researchers to rely on slow, black-box optimization or complex reinforcement learning. PyJama bridges this gap by providing a TensorFlow-based library that integrates seamlessly with NVIDIA Sionna.
The PyJama Framework
PyJama extends the NVIDIA Sionna ecosystem by introducing specialized blocks for transmitters (Tx), receivers (Rx), and jammers. Because every component in the pipeline is differentiable, researchers can apply Stochastic Gradient Descent (SGD) to "learn" optimal strategies for either side of the electronic warfare conflict.
The library supports a variety of jammer configurations, including barrage, pilot-targeted, and sparse jammers, operating in both time and frequency domains. On the defensive side, it implements advanced spatial filtering techniques:
- Projection onto Orthogonal Complement (POS): A technique that nulls interference by projecting the received signal away from the jammer's estimated subspace.
- Interference-Aware-Nulling LMMSE (IAN-LMMSE): A unified equalizer that treats jammer interference as spatially correlated noise to improve signal recovery.
Rather than relying on static datasets, PyJama leverages a simulation engine based on 3GPP Urban Micro (UMi) geometric channel models, ensuring that findings are grounded in realistic propagation environments.
Learning to Jam
A core capability of PyJama is the ability to "learn to jam" by optimizing power allocation over an OFDM resource grid. Using the Adam optimizer and custom loss functions—such as L1 loss for power efficiency or iteration-loss for coded systems—jammers can discover non-trivial strategies. For example, when targeting a single user, a learned jammer often identifies pilot symbols as the most vulnerable point, focusing energy there to corrupt the receiver's channel state information (CSI). This results in significant effectiveness gains; in certain scenarios, a learned jammer with -5 dB power can be as disruptive as a standard barrage jammer with -2.5 dB power.
Results and Discussion
Our simulations reveal critical vulnerabilities in standard mitigation techniques. While spatial filters like POS are effective against "dumb" jammers, they can be bypassed by intelligent adversaries. If a receiver uses specific "silent" slots in the resource grid to estimate the jammer's spatial signature, a smart jammer can learn to remain silent during those exact periods, rendering the receiver's estimate obsolete and the subsequent nulling ineffective.
Additionally, we investigated the impact of mobility and protocol compliance. We found that jammer mobility is often more detrimental than user mobility, as it rapidly outdates the spatial signatures required for nulling. Furthermore, jammers that violate the OFDM protocol (e.g., by omitting the Cyclic Prefix) create high-rank interference that is significantly harder to mitigate without also degrading the desired signal.
Conclusion
PyJama provides a powerful, open-source foundation for the next generation of wireless security research. By enabling differentiable programming in the context of electronic warfare, it opens the door for more sophisticated ML-based attacks and defenses. Future developments will focus on expanding time-domain jamming strategies and exploring competitive multi-agent scenarios where both jammers and receivers adapt in real-time.
Summary updated