Algorithm 1 The federated AMBR method. The K clients are indexed by k, rounds are indexed by r, n is the number of samples, $\left\{D_i\right\}$ represents the individual dataset owned by each client, E is the number of local epoch, B is the local batch size and w is the model parameters

Server executes: initialize $w_0$ for each round $r=1,2,\dots$ do     K ← (random subset of M clients)     for each client $k\in K$ in parallel do         $w_{r+1}^k$← ClientUpdate $(k,w_r)$     end for     $w_{r+1}\leftarrow {\sum }_{k=1}^K\frac{n_k}{n}{w}_{t+1}^k$ end for   ClientUpdate$(k,w)$://Run on client k Get parameters w from FL server for each local epoch i from 1 to E do     for batch $b\in B$ do         Extract audio embeddings $e_a$ and visual embeddings $e_v$ from local dataset $\left\{D_k\right\}$         Fuse the $e_a$, $e_v$ through the fusion network         Fine-tune and update w with loss function     end for end for return w to Server