Toward Robust EEG Lie Detection Systems: Multiband Source Space Dynamic Network States for Time Varying Deception

Human deceptive behavior is markedly time varying. However, most EEG liedetection systems still rely on static features and thus fail to accommodate distribution drift induced by evolving cognitive strategies, which limits system robustness. To address this issue,we propose a system oriented framework that estimates source space dynamic functional connectivity states across delta, theta, alpha, and beta bands and extracts temporal indices including occupancy, transitions, and dwell time.Across frequency bands, we observe a stable \u201cintegration-segregation\u201d architecture:integration states involve coordinated activity among high level control networks, whereas segregation states are dominated by sensorimotor coupling.  Guilty individuals show longer persistence in integration states, while innocent individuals preferentially occupy segregation states. We then fuse multiband global network features with state dynamic indices and evaluate deception versus honesty using repeated stratified cross validation. A random forest with feature selection achieves a mean AUC of 0.767 and performs significantly above chance in permutation testing. Stability analysis indicates that discriminative information is primarily concentrated in cross network coupling centered on the default mode network and in state dependent integration representations. These findings provide methodological support for building temporally sensitive and interpretable lie detection systems and lay a foundation for robust prediction in real world applications.