Single-channel jammer
Floods 433 MHz continuously at +17 dBm. Channel occupancy ≈ 100%. Receiver decodes nothing.
- PDR0.00
- SNR−35 dB
- SignatureTotal collapse
Edge AI · Imitation Learning · LoRa Physical Layer
When the radio gets jammed,
the network
heals itself.
SHARC is a TinyML system that listens for radio-frequency interference, classifies the attack pattern, and rewrites its own radio parameters — frequency, spreading factor, packet cadence — without a human, a server, or a cloud connection. It does this on a $5 microcontroller in under 10 milliseconds.
- Test accuracy
- 98.69%
- Model size
- 20.8KB
- Inference latency
- <10ms
- Channel states
- 4classes
01 / Problem
A radio you can silence for the price of a coffee.
LoRa is the radio backbone for millions of low-power IoT deployments — agriculture, utilities, conservation, healthcare in the field. Long range, multi-year batteries, license-free spectrum. It is also structurally undefended.
The physical layer relies on Chirp Spread Spectrum modulation. Two packets at the same spreading factor collide — and only one survives, whichever arrives stronger at the receiver. An attacker with a $5 module, a battery, and a known transmit power can drive the packet-delivery rate of an entire network to zero. Prior work has confirmed this with off-the-shelf hardware.
Worse: a reactive jammer doesn't need to be on the air all the time. It listens for the preamble, fires a short burst, and stays invisible 99% of the time. PDR ends up in the 0.2–0.5 range — degraded enough to break the network, intermittent enough to look like ordinary RF noise.
Channel-hopping jammer
Sweeps 8 sub-channels every ~4s. Each channel briefly clean, briefly hostile. Looks like ordinary multi-path on first glance.
- PDR0.70–0.90
- RSSI var700–1200
- SignatureHigh RSSI variance
Reactive jammer
Listens for ValidHeader IRQ, fires a 500ms burst, cools down 3s. Stealthy. Devastates targeted links.
- PDR0.20–0.50
- SNR var300–500
- SignatureIrregular inter-arrival
02 / Approach
Don't classify the attack. Imitate the expert.
Most jamming-detection papers stop at classification: "this is reactive jamming." That's useful information that nobody can act on. We took a different path.
Behavioural Cloning — a form of imitation learning — trains a small neural network not to label attacks, but to recommend the same response a domain expert would. The model maps eight RF observables directly to three binary actions: hop frequency, change spreading factor, randomise transmission interval. The expert policy gets baked into 20.8 kilobytes of TFLite weights.
Training is supervised and offline. Inference is on-device. No reward signal. No cloud. No backhaul. The model sees a feature vector, returns three bits, and a back-channel command frame carries those bits to the transmitter. Both nodes change parameters in the same packet cycle.
How the system fits togetherFIG. 01 · Behavioural Cloning · 8 → 32 → 64 → 32 → 3
The expert policy, in three rows.
Condition
freq_hop
sf_change
interval_rand
Why
Clean
0
0
0
No attack. Don't disrupt a working link.
Single-channel
1
0
0
Channel blocked. Vacate. SF change is wasted here.
Channel-hopping
1
1
0
Jammer follows freq. New SF changes collision profile.
Reactive
0
1
1
Preamble detector is SF-locked. Timing jitter breaks it.
03 / System
Three nodes. One closed loop.
Hardware
- MCUESP32-S3N16R8 · 240 MHz · 8 MB PSRAM
- RadioSX1278 Ra-02 · 433 MHz ISM
- SPISCK 13 · MISO 12 · MOSI 9 · CS 10
- PowerTargeting <3W operating
Radio config
- Centre freq433.0 MHz (adaptive 4 ch.)
- SF / BW / CRSF12 · 125 kHz · 4/5
- TX power+2 dBm (Tx) · +17 dBm (jammer)
- PDR window20 packets
ML pipeline
- ModelDense 8→32→64→32→3 · sigmoid
- LossBinary cross-entropy (multi-label)
- OptimiserAdam · lr 1e-3 · early stop
- DeployTFLite Micro · 20.8 KB
04 / Results
A jam, a detection, a recovery — in under five packets.
FIG. 02 · Packet Delivery Rate · attack & recovery cycle
PDR
Detection
Adaptation
Test accuracy
98.69%
across all three binary action outputs on the held-out test set
Clean
1.00 / 1.00 / 1.00
prec · rec · F1
Single-channel
1.00 / 1.00 / 1.00
prec · rec · F1
Channel-hopping
0.97 / 1.00 / 0.98
prec · rec · F1
Reactive
1.00 / 0.97 / 0.98
prec · rec · F1
- TFLite model size20.8 KB
- RAM during inference< 5 KB
- Inference latency< 10 ms
- Flash utilisation< 0.13% of 16 MB
- PSRAM headroom> 7.9 MB
- Training time (CPU)2–5 minutes
05 / Hardware
A real board. Not a render.
The custom PCB carries the ESP32-S3, the SX1278 Ra-02, AES-accelerated crypto, and a power stage targeting sustained operation under 3 watts. Drag to rotate. The geometry below is the actual model.
-
01
ESP32-S3N16R8
Dual-core Xtensa LX7 at 240 MHz with 8 MB PSRAM — enough headroom for the 64-unit hidden layer that standard ESP32 variants cannot accommodate in SRAM.
-
02
SX1278 Ra-02
433 MHz transceiver with full CSS modulation support, direct-register access for fine control over LDRO, CRC, and IRQ behaviour.
-
03
Hardware AES
On-chip cryptographic acceleration for AES — payload-layer protection is handled in dedicated silicon, not stolen from the ML inference budget.
-
04
Cost target
Built from off-the-shelf modules. A fraction of the cost of legacy frequency-hopping radios — and reprogrammable from a laptop.
Loading geometry…
06 / Applications
Where a jam is the difference between a number and a consequence.
Healthcare
Remote patient monitoring
Wearable ECG, SpO₂, glucose telemetry over LoRa in field clinics and rural hospitals. Medical equipment EMI looks a lot like reactive jamming — SHARC distinguishes the two and adapts SF or interval before an alarm is missed.
Defence & ISR
Battlefield sensor networks
Unattended seismic, acoustic, IR tripwires reporting to a forward operating base. Adversaries actively deploy jammers as precursor to incursion. SHARC's no-cloud, edge-resident inference is built for D3 RF environments.
Precision agriculture
Distributed soil & irrigation
Hundreds of sensors per gateway across hectares of farmland. Frequency-agile drone payloads and adjacent farms' gateways create exactly the channel-hopping profile in the training set. Hop + SF change recovers PDR.
Critical infrastructure
Pipeline & grid monitoring
Pressure, flow, valve, leak sensors across hundreds of km. RF disruption paired with physical sabotage is a real threat model. Multi-feature inference separates accidental industrial EMI from deliberate reactive attacks.
Conservation
Anti-poaching collar networks
GPS collars on endangered megafauna in remote reserves. Poaching crews use $50 jammers to blind ranger networks before incursion. Single-channel jam → instant frequency hop → animals stay visible.
Disaster response
SAR mesh communication
Portable LoRa nodes in earthquake or wildfire zones. The 433 MHz band becomes packed with amateur, emergency, and relief traffic. Looks like channel-hopping. SHARC moves the mesh to a clear sub-channel autonomously.
Smart cities
Utility & parking sensors
Dense urban LoRaWAN — thousands of meters per gateway. Customers occasionally jam their own smart meters to dispute bills. SHARC hops, recovers, and writes a tamper-evident log entry for the utility operator.
Autonomous systems
UAV / UGV telemetry
A reactive jammer on the telemetry link forces a lost-link failsafe and the mission ends without anyone being shot at. SF cycling defeats SF-locked preamble detectors; interval randomisation defeats timing prediction.
07 / Publication
The work, written down.
Research article
Evaluation of Low-Cost Jamming Attacks on the LoRa Physical Layer
Low power wide area network (LPWAN) technologies, particularly LoRa, are increasingly deployed around the world for applications such as smart agriculture, environmental monitoring, and industrial automation. Despite their advantages in long-range and low-power communication, the physical layer of LoRa remains vulne...
Read full articleConference paper · submitted
Behavioural Cloning-Based Anti-Jamming Action Recommendation for LoRa IoT Networks Using Embedded TinyML
An autonomous anti-jamming system for LoRa IoT networks employing Behavioural Cloning to recommend and execute targeted radio-parameter adaptations without human intervention. A compact multi-label BC model maps observed RF conditions directly to a three-dimensional binary action vector — converted to TensorFlow Lite and deployed as a 20.8 KB TinyML artefact on an ESP32-S3, achieving 98.69% test accuracy across four channel states.
Companion paper
Evaluation of Low-Cost Jamming Attacks on the LoRa Physical Layer
The empirical groundwork. An experimental evaluation of three jamming strategies against a LoRa link on ESP32 + SX1278 testbeds. Demonstrates that a $50 adversary can drive PDR to zero across all six spreading factors — motivating the need for the autonomous recommendation system above.
08 / Team
Four engineers. One supervisor. One radio.
MS
Mohammad Shayan
System Design · Firmware
TC-22038
AN
Syed Ali un Naqi Naqvi
ML Pipeline · Hardware
TC-22029
SR
Shahzaib Raza
Data & Evaluation
TC-22037
AH
Syed M. Aun Hasan
Research · Documentation
TC-22039
Supervisor
Dr. Aamir Zeb Shaikh
Associate Professor · Department of Telecommunications Engineering · NED University of Engineering & Technology
09 / Contact
Have a network worth protecting? Let's talk.
We're actively looking for partners with real RF environments — agriculture, conservation, utilities, infrastructure — to test the system in conditions we haven't seen in the lab.
If you have a deployment that needs resilience against interference (deliberate or otherwise), we'd like to hear what it looks like. Demonstrations, technical Q&A, hardware reviews — all on the table.
We aim to respond within 24 hours. No automated follow-ups.