Adarsh S Mcold start
000
resolving runtime01/05
← All work

Battlefield IoT

Edge ML threat detection for defence

Military IoT systems combining real-time analytics, edge computing and ML-based threat detection, deployed under DRDO and Army cybersecurity standards with a cross-functional team of ten-plus engineers.

Where
Centre for AI & Robotics, DRDO
When
2023 — 2026
Role
Project Engineer — systems design, team lead
Stack
Edge AI · Real-time analytics · ML threat detection · Embedded systems · IS 15500
40%Battlefield intelligenceimprovement via ML threat detection
100%ComplianceDRDO & Army cybersecurity and data privacy standards
+10Engineers ledcross-functional team

Problem

Battlefield sensing generates far more signal than a link back to a command centre can carry, and the moments that matter are exactly the moments when that link is least reliable. Intelligence has to be produced where the sensor is, not where the datacentre is.

Constraints

  • 01Inference at the edge — connectivity cannot be assumed
  • 02Full compliance with DRDO and Army cybersecurity and data privacy standards
  • 03IS 15500 conformance for reliability in military applications
  • 04A cross-functional team of 10+ engineers working to a single deployment

Architecture

SENSEEDGEDETECTASSURECOMMANDField sensorsdistributedEdge gatewaylocal computeReal-time analyticsstreamingML threat detectionon-device inferenceSensor fusionmulti-sourceCybersecurity controlsDRDO & Army standardsIS 15500 conformancereliabilityCommand picturebattlefield intelligence
fig.01Detection runs at the edge so that a degraded link downgrades the reporting, not the sensing.

Key decisions & trade-offs

  • Detection at the edge, not at the command centre

    Pushing ML inference onto the gateway means a sensor keeps producing intelligence when the uplink degrades. Backhauling raw telemetry would have made the most contested moments the least observable ones.

    Trade-offEdge hardware bounds model size and power budget, and updating a fielded model is a logistics problem rather than a deploy. Availability under degradation was worth both.

  • Compliance designed in, not audited in

    Cybersecurity and data privacy standards were treated as design inputs alongside latency and accuracy. In a defence deployment a system that fails accreditation does not ship, no matter how well it performs.

    Trade-offIt constrains the toolchain, the libraries and the data handling from day one, and slows early iteration considerably. The alternative is a rebuild at the end.

  • One deployment target for a ten-person team

    With 10+ engineers across disciplines, the main risk is divergence. Holding everyone to a single integrated deployment rather than parallel subsystem milestones kept integration continuous instead of catastrophic.

    Trade-offIndividual workstreams move slower and there is more coordination overhead per engineer. It is what made a single successful deployment possible.

Results

40%Battlefield intelligenceenhancement via ML-based threat detection
100%Standards complianceDRDO & Army cybersecurity and data privacy
+10Engineerscross-functional team led
1Deployed systemAI-driven Military IoT, successfully fielded