FAULT TOLERANCE AND GRACEFUL DEGRADATION OF CASCADE SOFTWARE FOR DETECTING FALSIFIED DIGITAL INFORMATION

The problem of keeping cascade software for detecting falsified digital information operational when its top tier, based on a large language model, fails is considered. Four classes of failure of this tier are distinguished: unavailability, overload, excessive latency and invalid output. A circuit breaker with closed, open and half-open states is proposed to detect them, taking into account both errors and response time. Five operating modes of the cascade are proposed: normal operation, deferred verification, decision by the lower tier with two thresholds, referral of uncertain messages to moderators and a safe state defined for each category of content. A criterion of minimum expected loss is proposed for choosing a mode. It takes into account the cost of a missed falsification, of wrongly blocked content and of a delayed decision under a constraint on the throughput of moderators. For the deferred verification mode, estimates of the backlog and of the recovery time after the top tier returns are obtained. It is shown that the degradation mode should be chosen separately for each category of content and verified by controlled fault injection.