Fabrice Kusters | Searcharxiv

arXiv2025

MEAL: A Benchmark for Continual Multi-Agent Reinforcement Learning

Benchmarks play a central role in reinforcement learning (RL) research, yet their computational constraints often shape what is studied. Despite the motivation of lifelong learning, most continual RL papers consider only 3-10 sequential tasks, as CPU-bound environments make longer sequences impractical. Meanwhile, continual learning in cooperative multi-agent settings remains largely unexplored. To address these gaps, we introduce MEAL (Multi-agent Environments for Adaptive Learning), the first benchmark for continual multi-agent RL. By leveraging JAX and GPU acceleration, MEAL enables training on sequences of 100 tasks in a few hours on a single GPU. We find that long task sequences reveal failure modes that do not appear at smaller scales.

cs.AI↗

arXiv2023

Jamming Detection in Low-BER Mobile Indoor Scenarios via Deep Learning

The current state of the art on jamming detection relies on link-layer metrics. A few examples are the bit-error-rate (BER), the packet delivery ratio, the throughput, and the increase in the signal-to-noise ratio (SNR). As a result, these techniques can only detect jamming \emph{ex-post}, i.e., once the attack has already taken down the communication link. These solutions are unfit for mobile devices, e.g., drones, which might lose the connection to the remote controller, being unable to predict the attack. Our solution is rooted in the idea that a drone unknowingly flying toward a jammed area is experiencing an increasing effect of the jamming, e.g., in terms of BER and SNR. Therefore, drones might use the above-mentioned phenomenon to detect jamming before the decrease of the BER and the increase of the SNR completely disrupt the communication link. Such an approach would allow drones and their pilots to make informed decisions and maintain complete control of navigation, enhancing security and safety. This paper proposes Bloodhound+, a solution for jamming detection on mobile devices in low-BER regimes. Our approach analyzes raw physical-layer information (I-Q samples) acquired from the wireless channel. We assemble this information into grayscale images and use sparse autoencoders to detect image anomalies caused by jamming attacks. To test our solution against a wide set of configurations, we acquired a large dataset of indoor measurements using multiple hardware, jamming strategies, and communication parameters. Our results indicate that Bloodhound+ can detect indoor jamming up to 20 meters from the jamming source at the minimum available relative jamming power, with a minimum accuracy of 99.7\%. Our solution is also robust to various sampling rates adopted by the jammer and to the type of signal used for jamming.

cs.CR↗