Where does poisoning collide?

In heterogeneous catalysis, poisoning is one route by which a catalyst deactivates: a chemical species binds strongly, often irreversibly, to the active sites on the catalyst surface and blocks the reactant molecules that would otherwise be converted there. Even trace amounts of a poison such as sulphur, lead, or chlorine can suppress a large fraction of activity, because a single adsorbed poison atom can deny access to several neighbouring sites. Bartholomew's 2001 review Mechanisms of catalyst deactivation sets poisoning alongside fouling, sintering, and volatilisation as distinct deactivation pathways. The referent here is the loss of a solid catalyst's working capacity, not harm to an organism or corruption of a learned model.[1]

Unit

Fraction of active sites blocked, read as the decline in catalytic activity over time

Recovery means

If the poison binds reversibly, activity returns on regeneration (oxidative or reductive treatment); strongly chemisorbed poisons deactivate the catalyst permanently

Measured by

Tracked by the fall in conversion or turnover frequency with time on stream, and by surface analysis of the adsorbed poison

In toxicology, poisoning is the injury a chemical causes to a living organism once the dose, route, and duration of exposure overwhelm the body's capacity to tolerate or clear it. The discipline rests on the dose-response relationship: the same substance may be harmless or therapeutic at a low dose and lethal at a high one, so that toxicity is a property of the exposure, not merely of the agent. Klaassen's Casarett and Doull's Toxicology: The Basic Science of Poisons is the standard reference for how poisons are absorbed, distributed, act, and are measured. The referent is damage to a biological system from chemical exposure, unrelated to a blocked catalyst or a corrupted training set.[2]

Unit

Severity of harm as a function of dose, route, and exposure duration

Recovery means

Many poisonings resolve with elimination, supportive care, or a specific antidote; others leave lasting injury or are fatal

Measured by

Characterised by dose-response endpoints such as the LD50, threshold and no-observed-adverse-effect levels, and biomarkers of exposure and effect

In machine learning security, poisoning is an attack carried out during training rather than at prediction time: an adversary who can inject or alter a fraction of the training examples steers the model that is learned from them, degrading its overall accuracy or planting a targeted error the attacker can later exploit. Because models are increasingly trained on data gathered from open or untrusted sources, even a small share of crafted points can shift the fitted parameters. Jagielski and colleagues' 2018 Manipulating Machine Learning: Poisoning Attacks and Countermeasures for Regression Learning formalises such attacks and defences for regression. The referent is the corruption of a learned model through its training data, a different thing entirely from a poisoned catalyst or a poisoned body.[3]

Unit

Model degradation or attack success as a function of the fraction of training data poisoned

Recovery means

Mitigated by data sanitisation, robust or certified training, and outlier removal rather than reversed after the fact

Measured by

Assessed by the rise in test error, or the targeted attack's success rate, against the proportion of poisoned training points