Decentralized secure aggregation (DSA) considers a fully-connected network of $K$ users, where each pair of users can communicate bidirectionally over an error-free channel. Each user holds a private input, and the goal is for each user to compute the sum of all inputs without revealing any additional information, even in the presence of collusion among up to $T$ users. Traditional DSA typically requires large key sizes to protect all information except for the input sum and the information of colluding users. To mitigate the source keys overhead, we study decentralized secure aggregation with arbitrary collusion and heterogeneous security constraints. In this setting, the inputs of a predefined collection of user subsets, called the \emph{security set} $\bm{\mathcal{S}}$, must be protected from another predefined collection, the \emph{collusion set} $\bm{\mathcal{T}}$. For an arbitrary security set $\mathcal{S}\in \bm{\mathcal{S}}$ and an arbitrary collusion set $\mathcal{T}\in \bm{\mathcal{T}}$, we characterize the optimal communication and source key rates. A key contribution of this work is the characterization of the optimal source key rate, i.e., the minimum number of key bits per input bit that must be shared among users for decentralized secure aggregation with arbitrary collusion and heterogeneous security constraints to be feasible. In general, this characterization reduces to solving a linear program.