Detecting Coordinated Inauthentic Behavior Under Symmetry Breaking: An Adaptive Memory-Guided Causal Framework

Detecting coordinated inauthentic behavior on social media remains a critical challenge, as many existing methods rely on correlation-based heuristics, fixed configurations, and heavy manual annotation. From the perspective of symmetry and asymmetry, coordinated campaigns often exhibit repeatable temporal and structural invariances (e.g., synchronized bursts and stable influence motifs), whereas adversarial adaptation and noisy environments introduce symmetry breaking and context-dependent deviations. To address this issue, we propose Adaptive Causal Coordination Detection (ACCD), a three-stage progressive framework with memory-guided adaptation. In Stage 1, ACCD introduces an adaptive Convergent Cross Mapping (CCM) module that learns embedding parameters across scenarios to recover invariant causal dependencies. In Stage 2, ACCD integrates active learning with semi-supervised classification to reduce labeling effort while preserving robust discrimination under asymmetric user behaviors. In Stage 3, ACCD employs an experience-driven validation module to self-verify detection results and mitigate spurious correlations across varying contexts. We evaluate ACCD on real-world benchmarks, including the Twitter IRA dataset, Reddit coordination traces, and TwiBot-20. Experimental results show that ACCD achieves an F1-score of 87.3% on coordinated attack detection, outperforming the strongest baseline by 15.2%, while reducing manual annotation by 68% and delivering a 2.8× speedup via hierarchical clustering optimization. Overall, ACCD 19 provides an accurate and scalable end-to-end solution that explicitly leverages symmetry (invariant coordination signatures) and asymmetry (evolving adversarial behaviors) for practical coordination detection.