Skip to main content
Back to timeline
arXivSource publication:

Defenders poison the eavesdropped signal: gradient perturbations inflate an attacker's system-identification error while touching only a quarter of transient samples

Synopsis

Against adversaries that identify a plant with a neural network and then act covertly, a defender injects small perturbations into the measurements the attacker eavesdrops, in decoy settings, so the learned plant model loses accuracy. The perturbations aggregate input gradients from several independently trained models and are confined to transients after setpoint changes. On a simulated DC-motor networked control benchmark, at equal root-mean-square magnitude they produce larger maximum absolute error than Gaussian and pseudo-random binary sequence poisoning, with the gap generally widening as magnitude grows; keeping only 25% of transient samples remains effective.

Source-provided article image: Gradient-Based Defensive Poisoning Against System Identification Attacks on Cyber-Physical Systems
arXiv

Interpretation

Rather than relying on passive detection alone, the defender actively poisons the data an attacker uses to train an identification model, degrading the neural-network-based identification and thereby the model accuracy that later covert attacks depend on. Earlier comparable protections either used switched control to hinder closed-loop model identification or used Gaussian input perturbations to provide differential-privacy guarantees; this instead targets the plant itself with gradient-directed poisoning signals. On a simulated DC-motor networked control system, maximum absolute error was tallied over models from 30 independent training runs; at equal root-mean-square magnitude the gradient poisoning's mean maximum absolute error was consistently above Gaussian and pseudo-random binary sequence poisoning, and the difference generally grew with perturbation magnitude.

The defender first trains a set of identification models on the same dataset with different random seeds, then computes, for each transient regime, the gradient of prediction error with respect to the input, accumulates it per sample and averages across models to obtain one generalized perturbation direction. The gradient indicates the input change that raises prediction error; after normalization to a target magnitude, only the components with the largest absolute values are kept and the rest set to zero, concentrating the perturbation on a few influential samples. Twelve perturbation profiles were evaluated, combining four magnitudes with three retention percentages; poisoning remained effective even when only 25% of transient samples were modified, indicating a small cost to sparsification.

Perspective

The evaluation runs on a simulated DC-motor networked control system, using ten passive-eavesdropping recordings of 60 seconds and 3001 samples each; the method targets decoy environments such as honeynets rather than operational systems, where perturbations could affect the physical process.

The evaluation assumes the defender's surrogate models and the attacker's models share architecture, input ordering and training procedure; stealth is measured only by perturbation root-mean-square and the number of modified samples, and sparse, deterministic perturbations aligned with setpoint transitions may be detectable by structure-aware analyses; the results measure prediction error of poisoned models and do not directly assess the detectability or operational impact of an attack derived from them.