Phase-Dynamic Model of User Interactions for Protecting Recommender Systems from Poisoning Attacks

Poisoning and shilling attacks remain a serious threat to recommender systems, especially as attackers increasingly mimic plausible profile statistics. This paper proposes an architecture-independent behavioral detection layer that models user interactions as short-window phase-dynamic trajectories rather than static aggregates. Interaction logs are transformed into temporal signals, reconstructed in phase space by delay embedding, and summarized by a compact 15-dimensional portrait combining recurrence-based, entropy-based, spectral, and stabilizing statistical descriptors. In a controlled targeted injection protocol evaluated over 10 independent runs, the statistical baseline achieved PR-AUC = 0.723 ± 0.037 and TPR@1%FPR = 0.029 ± 0.006, the dynamic block achieved PR-AUC = 0.831 ± 0.011 and TPR@1%FPR = 0.220 ± 0.050, and the full portrait achieved PR-AUC = 0.872 ± 0.017 and TPR@1%FPR = 0.291 ± 0.043. Sensitivity analysis showed that recurrence-only descriptors were parameter-sensitive, whereas the extended dynamic block formed a stable high-performance region across a broad range of embedding settings. An IQR-normalized aggregated risk score further demonstrated clear post-window regime separation during injection periods. The results indicate that poisoning attacks primarily deform the temporal organization of behavior rather than only first-order statistics. The proposed phase-dynamic portrait is therefore best interpreted as a complementary behavioral risk-scoring layer for auditing, filtering, and monitoring rather than as a standalone defense.