Detecting Feedback-path Delay Injection Attacks Using Interacting Multiple Model Filtering

What the math gives to ML

The paper offers a constructive Bayesian mechanism for detecting unknown, time-varying feedback delays: represent each candidate delay as a mode, run one Kalman filter per mode, mix their state estimates using Markov transition probabilities, and update mode probabilities from Gaussian observation likelihoods. Its transferable asset is the IMM architecture, which produces an online posterior over discrete hypotheses and an explicit alarm threshold. A neural-network analogue can monitor stale or delayed parameter, gradient, or hidden-state feedback in recurrent, distributed, or federated training by treating delay length as a latent mode. The key falsifiable prediction is that, after a persistent delay injection, the posterior of the correct delay mode grows approximately exponentially, with log-posterior slope determined by the KL separation between innovation distributions.