← Lab /Systems

Covert attacks can fully identify plant behavior while staying undetected

Systems2026-10-23T14:06:32Z Method · Simulation · arXiv Threat model

The claim

A MITM adversary can iteratively identify the full behavioral subspace of a closed-loop system and launch stealthy attacks using only intercepted I/O data

Victor Mattos, Amit Bhaya, Daniel Sadoc Menasche, Alan O. de Sá · arXiv Paper PDF

Published as: “Iterative Behavioral Identification for Covert Misappropriation Attacks”

So what

The work demonstrates that DeePC can be repurposed for covert attacks, highlighting a new class of threats that existing simple IDS schemes may miss.

The five-minute version

The paper shows that a MITM adversary can achieve full behavioral rank identification and launch a stealthy misappropriation attack on a closed-loop LTI system using a three-phase, model-free pipeline, with analytical stealth guarantees in the probing and attack phases and numerical support during iterative identification. Evidence comes from a simulation on a second-order DC motor where the attacker reaches rank n+mL in 23 steps and tracks a sinusoidal reference while the masked residual never exceeds the IDS threshold, though the unmasked residual does. The claim rests on noise-free simulations only.

What they found 1

Stealth is analytically guaranteed in the probing and attack phases, and numerically supported during iterative identification.
“Stealth is guaranteed analytically in Phases 1 and 3 and is numerically supported in Phase 2.”

What undercuts it the paper’s own results

Stealth is only numerically supported in Phase 2, lacking an analytical guarantee for that phase.
All analysis assumes noise-free LTI dynamics, so the guarantees may not hold under realistic noise.

Stated limits 3

All analysis assumes noise-free LTI dynamics.
the infinity-norm IDS itself is analytically tractable but weaker than CUSUM or χ2 tests; adapting the masking conditions to statistical detectors would require probabilistic arguments beyond the present scope.
Phase 2 costs O((qL)2) per step; Phase 3 solves the convex QP (23), scaling with NK and N. Both suit offline or slow-sampled SCADA use; real-time settings may need warm-starting.

What’s new here 4

Combines model-free behavioral system identification with stealth-constrained active probing in a closed-loop setting.
Achieves full behavioral rank with the minimal sample count T* = n+(m+1)L-1, matching open-loop bounds despite operating under a MITM constraint.
Provides analytical stealth guarantees for the probing and attack synthesis phases, and proves monotonic decrease of projection error to zero.
Integrates a DeePC-style convex QP to synthesize covert attacks that satisfy explicit infinity-norm IDS constraints.

Threat model stated by the source

AdversaryMan-in-the-middle positioned between plant and controller, intercepting and modifying both actuator and sensor channels.

CapabilityInject arbitrary additive actuator perturbations ua,t and apply sensor masking ya,t; full intercept history; knows IDS threshold, steady-state estimates, and peak-gain estimate.

AssumesPlant and controller are linear time-invariant and closed-loop stable; Dynamics are noise-free; IDS monitors only the masked output with an infinity-norm threshold

Artifacts & methods 6

Rademacher probing signalprimitive
Iterative rank-increasing experiment designprimitive
Behavioral subspace identification via Hankel matricessystem
Masking predictor (eq. 18)primitive
DeePC-style convex QP attack synthesissystem
Simulation of second-order DC motor with PI controldataset

Topics

behavioral systems theoryHankel matrix rank identificationDeePCcovert misappropriation attacksinfinity-norm intrusion detectionpersistent excitationmodel-free controlonline experiment design

← Previous Identity amortization lets Sybil attackers cut cost per identity up to 4.3×