API distillation campaigns try to hide model-stealing traffic by splitting requests across thousands of load-balanced accounts, which standard content classifiers often miss. However, request timestamps, specifically arrival time and stream completion time, reveal the underlying dispatching strategy, such as round-robin counters or key-shuffling pools. Using a 20 percent seed set, this timing analysis can forensically rank the remaining accounts of a campaign with 0.92 to 1.00 AUC within 20 minutes, even across thousands of real production accounts. While adaptive attackers struggle to hide rotating dispatchers using jitter, pinning keys successfully evades detection, though it can severely throttle throughput when quotas bind. Ultimately, timing exposes shared infrastructure patterns even when content filters fail.
Figures & tables
Fig. 1: High-pass occupancy correlation between the 40 keys of a pool (Baseline run). (a) LiteLLM shuffle: uniform weak turn-taking. (b) Shared counter with keys in arbitrary order. (c) The same keys ordered by circular seriation of (b) alone: the travelling wave of Proposition 1 .
Fig. 2: Correlation of keys at ring distance d (Baseline run). Dots: observed. Lines: Proposition 1 from the pool’s own visit delays and durations, with no fitted parameter.
Fig. 3: Eigenvalues of the hollowed key-by-key correlation matrix over the 703 keys of the Repeat run, against a time-shifted surrogate. Each outlier is a degenerate pair belonging to a rotation ring; the counter + jitter ring has two (Corollary 1 ).
Fig. 4: Cost of evasion (Evasion run, shared counter over 20 keys, 8 workers). Each point is one jitter level; detection z over 36 minutes against the throughput each worker keeps.