Jamsa: A Utility Optimal Contextual Online Learning Framework for Anti-Jamming Wireless Scheduling Under Reactive Jamming Attack

Abstract

Reactive jamming attack, enabled by programmable software defined radio, has become a serious threat to the ubiquitous wireless systems. Wireless scheduling under reactive jamming attack is important yet challenging. In this article, we consider the utility optimal scheduling problem for multi-hop wireless networks under reactive jamming attack. We propose pilot-aided channel probing , defensive power allocations, and jamming-aware route-selection mechanisms to counteract reactive jamming. By incorporating the link-level packet delivery ratio (PDR), we devise effective online learning approaches to improve network performance over time. The constantly sought optimal scheduling policy on the conflict graph is found to be between the well-known “experts” setting and the “Multi-armed bandits (MAB)” setting. More importantly, the PDRs treated as “expert advices” (contextual information) to recommend optimal schedules online over time makes the scheduling an intrinsic combinatorial contextual MAB (CMAB) problem. The utility optimal scheduling problem is formulated as a stochastic optimization framework to minimize the incurred regret. We define two rigorous reactive jamming attack models. An anti-jamming protocol is then proposed with provable utility and regret guarantees. We conduct both synthetic and real experiments to validate our theory. Experimental results show that our algorithm achieves 35%∼97% and 79%∼138% improvements of network delay and 23%∼58% and 67%∼108% improvement of network throughput over other state-of-the-art approaches for oblivious jamming and adaptive jamming, respectively. These results indicate that our framework could sustain high throughput under powerful reactive jamming attacks.