UTD mark
Md Ashikul Haque
Research

My research asks how a low-power wireless system stays correct and on time when the resource it depends on is contested. That resource may be spectrum taken by a jammer, airtime taken by a neighboring operator, or a token budget that runs out before a control deadline. Most of this work targets low-power wide-area networks (LPWANs), which connect battery-powered sensors to a gateway across several kilometers. In particular it targets LoRa, the radio technology that dominates such deployments.

Defense against jamming and coexistence with other networks are very challenging in LPWANs. Solutions designed for Wi-Fi, cellular, and other wireless networks cannot be adopted directly, because the Internet of Things (IoT) devices in LPWANs have limited computation, low transmission power, and batteries that must last for years. My dissertation, Dependable Low-Power Wide-Area Networking under Contention: Jamming, Coexistence, and Deadlines, addresses these challenges in four parts.

Jamming

An attacker can silence a whole deployment by transmitting noise on the channel the sensors use. The sensors are too limited in energy to detect the attack, and far too limited to answer it. I therefore build defenses that run at the gateway and decode packets while the attack is still in progress.

I published the first anti-jamming work for LPWANs, which assigns channels and transmission power through a game between the network and the jammer, and it improves packet reception by up to 31.83× (EWSN ’23). Every design since then has handled a stronger attacker, and none of them changes anything at the node. Gateway-side decoding recovers LoRa packets under a single jammer and improves packet reception by 83.91× (IoTDI ’24). A defense against jammers that transmit LoRa chirps rather than noise improves it by 106.56× (MobiHoc ’25). A defense against jammers that collaborate improves it by 100× (EWSN ’26).

The state-of-the-art work before this line addressed jamming by positioning at least three gateways in a line, and it was effective only against a single jammer. My designs impose no such placement and no overhead at the LoRa nodes, so they apply to networks that are already deployed.

Coexistence

LPWAN deployments now share unlicensed spectrum with one another, and a node has no energy budget to spend on negotiating for it. I moved the learning off the nodes and onto the LoRa Network Server, which assigns transmission parameters across a dense deployment with deep reinforcement learning. Against the decentralized state of the art this improves packet reception by 70.97% and cuts energy consumption by 62.91% (INFOCOM ’25). A follow-on line of work coordinates several operators whose controllers all learn at the same time, without asking any of them to disclose their traffic (ICLR ’26 workshop).

Deadlines

A control loop is useful only if its messages arrive before their deadlines, and contention is what makes that hard. Burst-MAC handles the traffic bursts that follow an event in a LoRa network, and it improves packet reception by 4.5× during a burst (RTSS ’24). ReMix reclaims airtime that a mixed-criticality schedule reserves and never uses, which raises high-criticality deadline satisfaction in industrial wireless sensor-actuator networks from 32.9% to 85.9% (ICCCN ’26).

The same question now reaches large language model (LLM) agents. An agent that drives a physical actuator inherits the deadline of the control loop it sits in, and generating tokens takes time that the deadline may not allow. Treating such an agent as a best-effort service is therefore not sufficient. Deadline-aware token scheduling cuts the critical failure rate of an LLM serving system from 92.5% to 23.8% (RTSS ’26).

Reach

Long Range Frequency Hopping Spread Spectrum (LR-FHSS) carries an uplink for tens of kilometers, including an uplink to low Earth orbit, but a LoRa downlink fails beyond 1 to 2 km. The nodes in that gap receive no acknowledgment and no control. I repurpose the LR-FHSS physical layer as a waveform synthesizer that produces LoRa-compatible chirps, with receiver-side reconstruction that corrects hop discontinuity, offset drift, and phase distortion. A LoRa receiver decodes these transmissions with no change to its RF front end, and reliability at long range improves by up to 79% (SenSys ’26).

Where this goes next

My goal is to make IoT and physical AI dependable by combining AI with next-generation wireless and real-time systems. Physical AI refers to AI that senses the physical world and acts on it through machines such as vehicles, robots, and drones. My doctoral research asked how a wireless network stays reliable and on time when jammers, neighboring networks, and competing deadlines contend for its resources. Physical AI and IoT beyond cellular coverage raise this question again with shared edge servers, short satellite passes, and moving vehicles. I will address it in multiple new projects.

A warehouse robot may need a stop command from an AI model within 100 to 250 ms. Serving systems for robot fleets raise the share of on-time actions on a shared graphics processing unit (GPU), but none bounds the probability that an action reaches its robot late. This project treats each action as a real-time task that spans a shared GPU and a wireless link, with ground robots as the testbed. I will schedule inference and transmission together to bound the probability that an action misses its end-to-end deadline. Agents that call tools pause, and the server must keep or rebuild their GPU memory. I will treat these pauses as self-suspensions and decide which memory to keep through a response-time analysis for chains of calls. Because timing assumptions can fail, I will build monitors that switch a machine to a verified safe controller.

Direct-to-satellite (DtS) IoT lets remote LoRa sensors reach a satellite in low Earth orbit. These sensors can warn of floods, wildfires, and other hazards, which leave little time to act. In July 2025, Texas’s Guadalupe River rose 32.5 feet in 1.5 hours. Jamming raises challenges that are new or harder with satellites. A satellite at 550 km passes over in about 10 minutes, so a lost warning may wait hours. One ground jammer can disrupt reception across a satellite’s whole coverage area. Because satellites are hard to update in orbit, a defense should need no change on the satellite. Acknowledgments are rare, so a device cannot tell jamming from collisions with unseen networks. Prior designs handle intermittent contact or jamming, but none certifies deadlines under both. Each device will derive its transmission times and channels from its LoRaWAN session key and frame counter. Devices already hold these keys, so this needs only a firmware update, not new key management. The network server will know when and on which channel each packet starts, but a jammer will not, so it cannot aim at a packet in advance. Given the predicted passes and a jammer with limited power, airtime, and reaction speed that knows everything except the key, I will certify which messages meet their deadlines and decline the rest. I will also develop AI models that estimate, from signal measurements across passes, whether missing packets were jammed, collided, or faded. The models will only choose among certified schedules, so a wrong estimate cannot break a certificate. I will test on a campus ground station and a drone gateway.

Any obstacle can hide an oncoming car or a crossing child from a vehicle’s cameras and lidars. Intersections account for about one-quarter of U.S. traffic deaths. Cooperative perception lets vehicles share detections, but only timely ones help. Standards set per-message latency and reliability targets, but no system bounds the probability that shared data arrives too late for a maneuver. Links can change within a single maneuver, so I will bound this probability over whole maneuvers and have a vehicle yield when shared data may be late. A vehicle should wait out congestion but rely on its onboard sensors under jamming, so I will develop models that tell the two apart from signal measurements at the receiver. When a jammer blocks reception at a vehicle, I will have roadside units that heard the shared detections forward them over the cellular network. Prototypes will run on software-defined radios and small vehicles.

Evaluation

I build and measure real systems. My experiments use software-defined radio (SDR) prototypes on USRP B200 with GNU Radio, outdoor deployments at kilometer scale, commercial off-the-shelf (COTS) LoRa nodes and gateways, a 15-node Tmote Sky testbed, and a three-car MentorPi testbed. At scale I use NS-3 and CARLA simulation, and PyTorch for the learning components.

The papers behind each part are listed on the publications page.