The implicit bias of gradient descent has long been considered the most likely mechanism that explains the superior generalization of over-parameterized neural networks without over-fitting, even when the training error is zero. Adversarial robustness is another key focus of research since it is crucial for the \textit{trustworthiness} of a model. A recent paper targets both generalization and adversarial robustness, concluding that implicit bias leads to non-robust solutions~\cite{frei2023double}. However, it does not account for the different behaviours of \textit{under-parameterized} and \textit{over-parameterized} neural networks concerning their implicit bias. In this paper, we dive into this trade-off, revealing those different behaviours through collaboration between consecutive layers. We quantify the collaboration between layers by a new concept named \textit{Co-correlation} which can be interpreted as \textit{alignment} of feature selections concerning maximizing outputs for each layer. We conducted extensive experiments that verified our proposed framework.