Department of Computer Science · Gulf University for Science and Technology
  1. Home
  2. Research
  3. Autonomous IoT Security

Towards Autonomous IoT Security

Automaton-based online vulnerability analysis and detection. Each IoT device is described by a model of its legitimate behavior, and the model is updated at runtime in response to conditions around the device.

Funder
Kuwait Foundation for the Advancement of Sciences (KFAS)
Duration
Sep 2026 – Sep 2028 (24 months)
Host
Gulf University for Science and Technology
Partners
University of Ottawa, Canada · Kuwait University, Kuwait
Overview

Protecting device functionality, not only data

IoT devices are now used in healthcare, energy, education and oil production in Kuwait, in line with the digital-transformation goals of Kuwait Vision 2035. Existing security measures mainly protect the confidentiality and integrity of data. Few address device functionality, that is, the risk that a sequence of legitimate-looking actions drives a device into an unsafe state. Static, rule-based defenses also adapt poorly to changing environments.

The project builds an adaptive security framework grounded in the theory of computation. Each IoT device is paired with an automaton that describes its legitimate behavior. An edge server uses observations of the environment and the system to relabel parts of each automaton with vulnerability scores. A threat can then be detected and contained as the device approaches a vulnerable state, before service is disrupted or the device is compromised.

Excessive blocking would make devices unusable. A reinforcement-learning controller therefore tunes the aggressiveness of relabeling to balance threat prevention against service availability.

Scope

Research questions

  1. Modeling behavior

    How can an IoT device's legitimate behavior be modeled as an automaton whose states, transitions and sequences can be relabeled as vulnerable at runtime, in response to environmental and system conditions?

  2. Scoring vulnerability

    How can a quantitative vulnerability score be defined and computed at runtime so that it captures proximity to unsafe states or sequences more expressively than a binary label?

  3. Adaptive control

    How can a reinforcement-learning controller tune the aggressiveness of relabeling to balance security against availability, and under what conditions does it converge to a stable, safe policy?

  4. Measurable benefit

    Does the integrated approach reduce exposure to vulnerable and unsafe states, relative to a static rule-based baseline, while preserving each device's declared critical functions?

Design

System architecture

The system comprises interconnected IoT devices, the wireless network that links them, and one automaton per device. An edge server collects observations, updates each automaton and reconfigures system parameters. Each operating period has five stages.

  1. ES

    Environment sensing

    Devices observe their surroundings and collect only the data required, to avoid overloading the network and the analysis.

  2. DT

    Data transfer

    Observations are sent over the wireless network to the edge server, with communication resources managed to avoid interruption.

  3. EP

    Edge processing

    The edge server maintains a compact model of the current environment and system state.

  4. AU

    Automata update

    States, transitions or sequences are relabeled as vulnerable, each with a score, based on new observations and history.

  5. RLT

    RL tuning

    Over longer horizons, an RL agent balances threat prevention against service availability.

Idle Measure Report Firmwareupdate Remotedisconnect v = 0.4 v = 0.8 start
An illustrative smart-meter automaton. Normal operation cycles through Idle, Measure and Report. When conditions change, the edge server assigns vulnerability scores (v) to states such as a firmware update or a remote disconnect, and execution paths approaching those states are flagged before the states are reached.

Relabeling and vulnerability scoring

Threats rarely correspond to a single state, so the relabeling algorithm must decide when to relabel, what to relabel (a state, a transition or a sequence) and how strongly. Detection is based on proximity: the closer the current execution is to a vulnerable part of the automaton, the higher its score. Candidate formulations include reachability distance and weighted risk propagation.

Security versus availability

Each state corresponds to a function the device provides, so every block has a cost. A security control parameter sets the share of states that may be flagged, and a reinforcement-learning agent learns this parameter from network conditions and threat history. If the agent does not converge, the system falls back to a bounded heuristic policy.

Testbed

Evaluation will use a laboratory testbed of programmable IoT nodes (Raspberry Pi devices with sensors, connected through a gateway) that emulate a smart meter, a sensing and actuation node and a network node. The University of Ottawa will provide access to the uOttawa–IBM Cyber Range for additional datasets and attack scenarios.

Targets

Key performance indicators

AreaIndicatorTarget
TechnicalVulnerability detection accuracy≥ 90%
EfficiencyAverage automaton update latency≤ 200 ms
AdaptabilityRL controller convergence< 10 epochs
ImpactMisuse incidents, downtime and manual interventions relative to a static rule-based baselineReduced
DisseminationPeer-reviewed papers≥ 4
OutreachWorkshops and seminars at GUST≥ 4
Outputs

Deliverables

  • Automata-based model of IoT device behavior (software module)
  • Reinforcement-learning adaptive defense module
  • Integrated system prototype, validated on the laboratory testbed
  • Three to four conference papers and one to two journal papers
  • Openly available reference implementation and dataset repository
  • Research assistant training and student workshops on IoT security
Plan

Timeline

The project runs for 24 months in six phases, each with acceptance criteria.

  1. Months 1–3 · Sep–Nov 2026Current

    Setup and requirements

    Finalize requirements, procure the testbed, arrange access to the uOttawa–IBM Cyber Range, specify the static baseline and onboard the research assistant.

  2. Months 4–8 · Dec 2026–Apr 2027

    Automaton modeling and device profiles

    Define the automaton formalism and model representative devices.

    Gate: at least three device automata, including a complete smart-meter model.

  3. Months 9–13 · May–Sep 2027

    Relabeling and vulnerability scoring

    Implement the triggering mechanism, the selection of relabeling targets and the scoring function.

    Gate: detection accuracy ≥ 90% and update latency ≤ 200 ms.

  4. Months 14–17 · Oct 2027–Jan 2028

    RL controller design and training

    Define states, actions and rewards; train the controller and integrate it with relabeling, including safe fallbacks.

  5. Months 18–21 · Feb–May 2028

    Integration and prototype

    Run the complete sensing, edge, relabeling and RL pipeline on the testbed.

  6. Months 22–24 · Jun–Aug 2028

    Evaluation and dissemination

    Benchmark against the baseline, report KPIs, publish results and release the code and datasets.