Wireless networks are increasingly vulnerable to sophisticated attacks that exploit predictable patterns in communication standards. Among the most effective of these threats is pilot jamming, where attackers target the special reference symbols (pilots) used in OFDM-based systems to estimate and equalize wireless channels. 

Research has shown that pilot jamming attacks can achieve 100% effectives and be more efficient than conventional jamming, making them a critical security concern for next-generation wireless networks.

Through the SoBigData.it Transnational Access (TNA) program, I had the opportunity to spend three weeks at the University of Catania, where my team from Northeastern University collaborated with leading experts in multi-armed bandit (MAB) optimization such as Profs. Sergio Palazzo, and Fabio Busacca and Raoul Raftopoulos to address this challenge. The result is PilotBandit, a novel framework that leverages intelligent learning algorithms to dynamically randomize pilot positions and evade jamming attacks.

The Challenge: Predictability in Wireless Standards

Modern wireless standards like WiFi and 5G use fixed and/or predictable pilot locations to ensure reliable communication. While this standardization is essential for interoperability, it creates a fundamental vulnerability: attackers can predict exactly where pilots will appear and concentrate their jamming power on these critical symbols. Our previous work at Northeastern using Deep Reinforcement Learning to reposition pilots in a pseudo-random fashion has showed promise with a 4x throughput improvement under attack, but suffered from poor sample efficiency, requiring weeks of training and resulting in predictable, exploitable behavior.

The Solution: Contextual Multi-Armed Bandits

During our collaboration at UniCT, we developed a more elegant and efficient approach using contextual bandits. Unlike traditional MAB algorithms that treat each decision independently, contextual bandits leverage real-time channel conditions and information on the attack as context variables to make informed decisions about optimal pilot placement. This approach offers several key advantages:

• Superior sample efficiency compared to deep reinforcement learning;
• Adaptive pilot positioning based on instantaneous channel state information and attack location;
• Balanced reward mechanisms that account for both short-term performance and long-term security; and
• Robustness against channel stochasticity and outliers.

What’s next?

Our team developed comprehensive system models integrating the MAB framework with OFDM pilot allocation mechanisms. We established VPN access to the Arena testbed at Northeastern University, a state-of-the-art wireless research facility spanning two floors with over 30 software-defined radios connected to 100 ceiling-mounted antennas. Arena's rich fading environment with time-varying channel characteristics provides the ideal testing ground for PilotBandit.

The experimental plan includes data collection across varying antenna configurations, radio settings, interference profiles, and channel conditions. We will use data analytics to identify clusters of optimal pilot configurations and train contextual MAB policies for real-time pilot randomization.

This project exemplifies the power of international collaboration in tackling complex security challenges. By combining Northeastern's expertise in wireless security and experimental systems with UniCT's deep knowledge of MAB optimization, we have laid the foundation for breakthrough advances in anti-jamming technology.

We are committed to open science principles and plan to release curated datasets to the research community and publish our findings in premier venues. This work will enable other researchers to build upon our approach and accelerate progress in wireless security.

The collaboration also strengthens research ties between Europe and the United States, paving the way for future joint funding proposals through programs like the NSF-MUR bilateral agreement.